Three native skills that run the Diolog Tasks issue pipeline — triage, plan, and worker — entirely inside your interactive Claude Code session (no Agent SDK, so usage draws from your standard interactive allowance, not the separate Agent SDK credit). Formerly linear-issue-pipeline; now targets the diolog-tasks MCP (mcp.diolog.com.au). tasks-triage runs a technical + Specification Sentinel product review and posts a non-technical readiness comment; tasks-plan classifies a tier, investigates the codebase, writes docs/plans/{id}.md and comments the repo-relative path (no file upload); tasks-worker implements a planned issue in an isolated worktree via dynamic ultracode workflows (understand → implement → rebase → acceptance review → resolve), with no remote PR. The plan and worker fan out parallel subagents via the Workflow tool to speed up large investigations. Recalibrated for current-model behaviour (Anthropic's Claude Opus 5 prompting guidance), ported from the generalized twin: tasks-worker's per-phase gate is explicitly a conformance check (evidence-shaped — the clause, the file:line, the gate you ran) rather than a second correctness review stacked on a phase that already reviewed itself; adversarial finding-verification is aimed at Criticals, structural fixes, and locked-decision reversals rather than blanket 1:1, and no reviewer is briefed to be conservative; the completion comment keeps to tables/counts/drops/gates rather than narrating the phases; and tasks-triage/tasks-plan gain the effort dial plus length budgets. All three point at the twin's new pipeline-wide canonical feature-spec-pipeline/skills/work/references/model-and-effort.md. v2.4.0 (from the WEB-4905 110-ticket audit): a fourth skill, tasks-verify — an independent fresh-session acceptance verifier that re-derives requirements from the ticket and closes them on typed behavioural evidence (browser measurements, exercised requests, stored-row counts) before a ticket is treated as done; evidence-typ
4 skills
Four native skills that run an issue-tracker-free, markdown-doc feature pipeline — /triage, /plan, /work, /gap-fix — entirely inside your interactive Claude Code session. /triage turns a feature idea (inline text or a notes file) into a versioned spec at docs/specs/spec-DIO-0001.md (id allocated from docs/feature-specs/LEDGER.md), running a codebase grounding pass plus a Specification Sentinel product/UX/compliance review and appending a non-technical readiness section. /plan classifies a tier, investigates the codebase, writes docs/plans/plan-DIO-0001.md and links it from the spec. /work implements the plan in an isolated worktree via dynamic ultracode workflows (understand → implement → rebase → acceptance review → resolve), appends a progress section to the spec, and leaves the branch local — no remote PR. /gap-fix runs after /work as the remediation step: it re-enters the work branch, ALWAYS re-audits the delivered code against the original spec (merging in any human/QA gaps you provide), and implements the fixes in code (file-disjoint fan-out + typecheck gates), looping audit→fix until only optional Low items remain, then commits locally and appends a gap-fix progress note — no remote PR. The plan, work, and gap-fix skills fan out parallel subagents via the Workflow tool (in small batches to avoid throttling). A doc-driven sibling of diolog-tasks-pipeline that needs no Diolog Tasks MCP. Recalibrated for current-model behaviour (Anthropic's Claude Opus 5 prompting guidance) with a new pipeline-wide canonical reference, skills/work/references/model-and-effort.md: the model × effort lane calibration (per-lane effort levels, `low` as the level built for subagents, effort-before-model because REVIEWER ≥ WRITER is about capability class not spend, `xhigh` + 64k `max_tokens` for long-horizon runners, hold-effort-constant for the prompt cache, never hardcode a dated model id in a wire self-check); the oracle-checks-vs-re-reading-your-own-work distinction that says whic
4 skills
Quality-assurance automation for the diolog-swe-bench benchmark corpus: judge whether a task is difficult enough and fairly verified (bench-task-judge), re-run the whole-corpus review with any model (bench-corpus-review), and author new gated tasks including ui design tasks (bench-task-author).
3 skills
Three skills that write publish-ready, on-brand content in Diolog's brand voice (Australian investor-communication SaaS; 'the workspace for everything investor-facing'; tagline 'Disclosure, without doubt.'), all driven by one bundled brand-voice persona and gated by one deterministic voice-lint. create-diolog-article writes long-form articles / blog posts / thought-leadership across the investor-education, IR-best-practice, thought-leadership and announcement registers. create-diolog-marketing-copy writes website marketing copy and marketing emails (benefit-first heroes, feature-to-'so what' bullets, problem-callback CTAs, subject/preview/body emails). create-diolog-business-case writes buyer-facing ROI / justification documents for the CFO / Company Secretary / board (status-quo cost -> what changes -> proof/security/compliance posture -> measurable outcomes -> next step). Every skill hard-encodes the house rules: no em or en dashes (spaced hyphen only), Australian English, sentence-case headings (except product feature names), every statistic carries an inline (Source, Year) with zero fabrication, compliance confidence capped at 95%, no financial advice or guaranteed outcomes, the Diolog narrative arc, correct register selection, and critique the old way of IR never the reader. The deterministic diolog_voice_lint.py hard-fails on em/en dashes and banned/AI-tell phrases and advises on American spellings, over-claim/guarantee language and citation coverage. The fuller brand counterpart to the deployed diolog-content-writer / diolog-linkedin-writer generation prompts; personal-voice LinkedIn posts remain with create-luke-content.
3 skills
Three macOS-native UI skills sharing one bundled knowledge corpus (the full macOS 26 'Tahoe'/Liquid Glass HIG library + a token-level DESIGN.md + the macOS 26 UI Extraction & Native Translation framework). `macosify` refits an existing UI to feel authentically AppKit-native on macOS 26, correcting iOS-derived looks (Mac Catalyst / 'Designed for iPad' on Apple Silicon) toward true native. `macos-ui-analyst` (the GlassReader persona) analyses a static macOS 26 screenshot into a structured Native Translation Report (AX-tree inventory + SwiftUI/AppKit mapping) and merges durable learnings into one accumulating knowledge file. `macos-storyboard` reads the entire corpus, then builds an interactive StyleX + Storybook storyboard of a specification's user flows — a reusable design system (tokens → elements → composites) with every screen rendered across happy / empty / loading / error / AI states, using the design-craft and frontend-design skills for bespoke, accessible direction.
3 skills
Convert PowerPoint (.pptx) ⇄ the lecturn.deck/1 JSON slide schema, and author/validate that JSON, with the bundled spec and a stdlib-only Python two-way converter. Import a .pptx to an editable deck/slide/element tree (media extracted, fidelity F0–F3 reported, no silent loss), or build a structurally valid .pptx from a deck JSON. Includes a specialist agent for multi-step convert-validate-verify loops.
1 skill · 1 agent
Turn a feature's requirements + the running app into an acceptance-criteria-traceable e2e suite in the project's own e2e harness — then go proactive: seven driven-and-asserted sweeps beyond the ACs (state-matrix forcing via interception, fault injection incl. offline/timeout/double-submit, interaction-integrity + reactivity, keyboard + axe a11y floor, data-shape stress over API-seeded edge fixtures, security surface incl. forged-action/IDOR/channel-authz, multi-user two-context realtime), a rendered-quality handoff that invokes the /design-review skill (deterministic WCAG/layout gates + judged craft/state/flow passes) and the mockup-fidelity layer where a design-of-record exists, surgical fixes for tractable bugs, and guard promotion — every sweep becomes a permanent CI/pre-push gate invoked by a named script, with new surfaces inheriting by enumeration. Coverage axes are enumerated and the sample declared up front; committed specs carry their run record. Lanes for native apps (Maestro deep-link-first, XCUITest, performAccessibilityAudit, and an Accessibility-tree runner for macOS apps XCUITest cannot reach - assert AXPress not just identifiers, fixture-mode determinism that also avoids stealing foreground focus, and the measured traps that make an AX harness blame the app), MCP servers (acceptance = a real MCP client driving the built server), and marketing sites (route x viewport loads, link/heading/meta audits, claim-grounding invariants). v1.4.0 phases 0-9, from the portfolio QA plan + the WEB-4905 audit. v1.6.0 moves the default runner off @playwright/test to node:test driving Obscura over CDP, wires the a11y floor by injecting axe-core (color-contrast lands in incomplete, never violations, under that engine), and states plainly what the conversion costs: no fixtures, no auto-retry, no parallel workers, no HTML reporter, no trace viewer.
1 skill
High-signal code review for NestJS, Next.js (App Router)/React 19, frontend HTML/CSS/React, and React Native. Depth levels (quick/standard/deep) scale token cost; area targeting (frontend, nest, next, mobile, paths) and focus lenses (bugs, security, perf, tests, components, a11y, dead-code, debt, deps, dx) scope each run; a token-light prepush gate scans the outgoing diff. Modular checklist routing, mitigating-controls auto-discovery, parallel sharding for large diffs, independent batched Sonnet verifier fan-out with five gates (API existence, versions, mitigation-elsewhere, proportionality, reachability), and cross-run learned suppressions. Quality lenses adapted from shadcn's improve skill (MIT); noise-reduction patterns from 2025-26 AppSec research (Corgea, Aikido, ZeroPath).
1 skill
Produce the company-overview markdown that the investor-portal generator reads: what a listed company does, who leads it, what it has lodged, and every image and PDF it publishes — crawled from the company's own site so every fact is attributable rather than authored. Pairs with design-md-from-website to form the two inputs generate-investor-portal requires. Carries the output contract the generator parses, including the exclusions (privacy policies, crawler scaffolding, cart and login pages) that have silently become fake business units on real runs.
1 skill
Write any content in Amy Benson's authentic voice (Diolog Founder & CEO), routed through eight persona variants: LinkedIn posts and thought leadership, long-form articles/blog posts, marketing content (announcements, launch/landing copy grounded in the diolog.app copy she curated and aligned with the Diolog voice & audience guide), sales/partnership outreach email (cold intros, consultative follow-ups, objection handling, gracious declines), Diolog investor updates, Slack/chat messages, short-form, and speaking notes for panels/podcasts/pitches. One evidence-anchored base voice (warm first, sharp underneath, direct questions with low-pressure asks, spaced-hyphen habit, no em dashes, AU spelling) synthesised from ~5,000 words of her verbatim email, LinkedIn, Slack and interview corpus (bundled, redacted), with a deterministic voice lint carrying her measured stylometric fingerprint. Generated with the create-voice-persona factory. v1.3 folds in Amy's own editorial review: broken-up LinkedIn formatting, genuine-question closes, peer-audience floor, referent discipline, register fencing for marketing, plain offer phrasing, no meta-labels, and retired phrases (bricks/Rome, 'came up through') enforced by the lint. v1.4 refreshes the bundled AI-signs field guide with the July 2026 empirical research layer and encodes the research-backed guards (plain copulas, epigram budget, hook-transition bans) in the base voice and lint config. v1.4.1 syncs the field guide's once-per-document epigram/contrast rule. v1.5 syncs the repeated-phrase lint check and the updated field guide (countable texture budgets, scaffold-slot trap). v1.5.1 field-guide sync. v1.5.2 lint pattern sync. v1.6 adds the meetings corpus: ~150k words of Amy's live sales/discovery calls (2025-26, six segments, redacted) distilled into references/research/amy-meetings-corpus.md - spoken hedge hierarchy corrected (kind of / I guess / to be honest dominate; I'd say / I find demoted to interview register), signature
1 skill
Build a high-quality self-contained HTML slide deck for a named company from three inputs — its company overview, its DESIGN.md design tokens, and a source document or prompt — assembling slides from a bundled library of proven layout templates and generating photography with the media-gen-pro MCP. Use whenever someone wants a deck, slides, a presentation, a quarterly or results deck, an investor update, a board pack or a strategy deck built for a specific company, especially when they hand over a company overview, brand tokens and an announcement or brief and ask for slides from them. Carries a generated theme block derived from the company's own type ramp by a single ratio, fifteen filled-and-numbered slide layouts, the figure-ledger and zero-based-axis discipline a disclosure-derived deck needs, media-gen-pro prompt and grade-lock recipes with the never-a-real-person rule, and a measured gate suite for four render failures that are invisible in a screenshot — stage clipping, print containing-block collapse, copy painted under a full-bleed image, and pinned chrome stretched by the wrapper-collapse guard. Supplies its own missing inputs — company-overview-from-website crawls the overview and design-md-from-website measures the DESIGN.md off the same site — so a company URL plus a source document is enough to start, and hands the built deck to design-review before completion. Carries a Diolog studio-runner target (references/eve-runner.md) so the same grounding governs decks authored as diolog.deck-source/1 deck.json by the Eve agent, where the HTML shell and scripts do not apply. Not for a .pptx or editable Office handoff (deck-craft's lecturn JSON target), an investor portal page (create-investor-portal-free), or a deck with no brand and no source material (deck-craft's direction round).
1 skill
Write any content in Diolog's company voice, routed through four register personas: long-form A4 guide copy (the master files create-diolog-guides renders), articles/blog posts (investor education, IR best practice, thought leadership, announcements), website + email marketing copy, and buyer-facing business cases. Successor to the diolog-brand-voice skills, rebuilt on the create-amy-content architecture: one evidence-anchored base voice (plain and literal beats clever, the skim test, epigram budgets, verbatim positioning beats, measured confidence shown never stated, spaced hyphen never em dashes, AU English) synthesised from the team voice guide, the live diolog.app copy, and Amy's line-by-line 2026-07 review of shipped AI-drafted long-form; a review ledger that compounds future corrections; the July 2026 empirical anti-AI-writing research layer; and a deterministic voice lint with per-register format budgets. v1.0.1 adds the stylometric fingerprint extracted from ~26k words of the live diolog.com.au corpus (marketing pages, non-legal help articles, all blog articles; AI-drafted resources excluded). v1.1 adds the revision protocol (item-by-item feedback disposition - applied/overruled/routed-and-landed - plus restatement and duplication sweeps) after the guide revision validated clean on copy rules but evaporated routed design feedback, left a stale spec restatement, and duplicated an epigram across a spread; ledger entries 10-12 record the classes. v1.2, from a blind generalisation test (clean-room generation on a fresh topic, class-blind adversarial review): seeded approved sample anchors with an approval-plus-clean-lint admission rule and an accretion loop; data-shaped colour counts as a claim; scaffold slots never filled with invented psychology; countable texture budgets replace vary-by-intention; lint gains the repeated-phrase check and repeat_allowlist; ledger entries 13-15. v1.3 separates rulebook from casebook: the review ledger is maintenance-only and ou
1 skill
Build a multi-page, print-ready A4 guide in Diolog's brand design system as one self-contained HTML file that scrolls as a document on screen, exports to a clean A4 PDF, and carries screen-only micro-interactions that never reach the print. Use whenever the deliverable runs longer than one page: a guide, briefing, playbook, handbook, whitepaper, explainer, report, or 'turn this markdown into something we can send investors'. Sibling of create-diolog-infographic, which lands one finding on one page; this one sustains an argument across twenty, which is a systems problem rather than a composition problem. Mandates a design layer before any page exists (a spacing scale, a column grid, one shared baseline, print-first type, motion tokens), shapes the document into three to five page archetypes rather than twenty one-offs, and gates each archetype through design-craft's per-unit critique before reproducing it. Renders approved company-voice copy verbatim and builds all imagery as live HTML and CSS rather than raster. Ships a zero-dependency Node CDP harness that measures ink rather than boxes: guide-qa (cap-height alignment via baseline probe and canvas TextMetrics, optical gutters, stretch voids, orphan wrap rows, spacing-scale adherence, cross-page baseline drift, clamped-box centre drift, Range-based widows, WCAG AA contrast with alpha blended up the ancestor chain, plus per-component crops at DPR 3), scope-lint (selectors that cross a component boundary, dead rules, colliding global classes), page-fit (A4 fit plus the collisions overflow:hidden conceals, footer overlap, computed font families), motion-check (print and reduced-motion resting-state safety, then mid-flight frame capture), and export-pdf (verifies the produced PDF's page count, sheet size, link annotations, and that no transient animation label leaked into the ink). Carries an explicit verification epistemics: a gate is downstream of the findings that motivated it, so a clean gate means no known defect i
1 skill
Create a print-ready, single-page A4 infographic in Diolog's brand design system to communicate one finding, comparison, or metric (a benchmark result, a cost or quality trade-off, a KPI, a 'we measured X' story). Themes everything from the live Diolog website design system (financial-broadsheet plus instrument-panel: Newsreader serif with one italic accent, Inter body, JetBrains Mono for tickers and data labels, navy ground used once, one blue accent, accent-bright for data on navy only, semantic colour always paired with a dot or sign, no em or en dashes, sentence case, no second hue, no emoji, no hype). Writes any copy in Luke's voice via create-luke-content and runs its voice lint. Embeds the three brand fonts inline as base64 so the page is self-contained for print and for publishing as an Artifact, and pulls the diolog droplet mark plus lowercase wordmark. Charts a finding honestly (equal-length tracks with differing fills for proportions, x versus y with both values shown and the winner styled, one dominant focal number, never a comparison of success or task counts) and verifies the page renders and fits exactly one A4 page (~794x1123px at 96dpi) by serving over http, rendering with a cache-busting URL, measuring the page height, confirming the fonts loaded, and trimming until it fits. Leans on design-craft and ux-craft for visual and UX craft. Use whenever someone wants an infographic, a one-pager, a shareable stat graphic, a poster of a result, or a designed chart of data in the Diolog look, even if they only say 'turn these numbers into something I can share'. Not for multi-page guides, briefings or whitepapers (create-diolog-guides), multi-slide decks (customer-deck-builder), email product mockups (email-mockups), or a full website or app UI (design-craft). v1.2 adds a mandatory Step 0 design layer (a spacing scale every value draws from, print-first type, a margin column sized to its widest glyph run, line-height 1 on display glyphs, optical nudges via t
1 skill
Turn a Diolog web-app disclosure-consistency report into a standalone, on-brand HTML artifact a Disclosure Committee can read, print and circulate before an announcement is released. Takes the saved HTML of the app's consistency-check page plus the subject company's DESIGN.md and produces one self-contained file: decision first, the items to resolve, an honest coverage bar, every tested statement, the supporting record and the anticipated questions. Use this whenever someone hands over a disclosure consistency report, a consistency check, a pre-release disclosure audit or an "AI compliance check" and wants it as a shareable or printable page, wants the app's report turned into a document, wants the Alfabs consistency audit rebuilt for another company or another quarter, or asks for a consistency report restyled in a company's brand - even if they only say "make this report look good" or "turn this export into something I can send the board". Treats the app HTML as the sole source of truth and invents nothing: no report IDs, owners, due dates or attributions the source does not carry. Reframes the app's misleading pass-rate as checked-versus-not-checked coverage, because a statement nobody could check has not been found inconsistent. Not for building an investor portal (use create-investor-portal-free) or for running the consistency analysis itself.
1 skill
Build a complete, single-page public investor portal for a listed company from its DESIGN.md design-token file and a company overview markdown file, branded entirely in that company's own tokens with a Diolog banner, app band and footer credit layered on top. Use this whenever someone wants an investor portal, investor page, investor relations page, IR site, shareholder page or "investor hub" built for a specific company, wants the Alfabs investor portal rebuilt for a different issuer, hands over a DESIGN.md plus a company overview and asks for a site or landing page from them, or asks for an investor-facing page that search engines and AI assistants can cite - even if they only say "make ACME an investor page from these two files" or "do what we did for Alfabs but for this company". Orders the page on research into what search-referred and LLM-referred investors actually need (dated sourced company facts first, disclosures second, share price demoted beneath both, governance as required ASX furniture), emits Corporation structured data and a crawler-permissive robots.txt, and carries the ASX Listing Rule 15.7, Listing Rule 4.10.3 and section 769C constraints that shape the build. Not for a disclosure-consistency audit page (use create-disclosure-consistency-page), a Diolog-brand marketing page (design-craft), or a DESIGN.md extracted from a site (design-md-from-website).
1 skill
Generates dense, agent-ready persona definitions from a target role — identity kernel, responsibility matrix with classification tags ([CRITICAL] / [WORKFLOW] / [POWER-USER] / [GOLDEN-NUGGET]), decision frameworks, maturity model, selection matrix, capability heat map, dependency graph, metrics, constraints, and two interaction examples. Tuned to current Claude model behaviour (literalism, adaptive thinking, positive framing). Evidence base from 2025 persona research (EmoCharacter, LABE, PPA, CPER, Over-Specification Paradox) ships in references/playbook.md.
1 skill
Build an interactive UI storyboard as a StyleX + Storybook project from a product specification and a user-supplied DESIGN.md. The DESIGN.md (the product's own brand/design system) drives the look — this skill is NOT locked to a macOS aesthetic and renders in whatever visual identity the brand specifies. It reads a bundled HIG library purely as the interaction, behaviour, structure, and accessibility backbone (layout, master–detail/list/table grammar, modality, feedback, focus/keyboard, motion, empty/loading/error/status, writing), applies the design-craft and frontend-design skills for distinctive, accessible, AI-slop-resistant direction, derives the user flows + screens, builds a reusable StyleX design system (tokens → elements → composites) from the DESIGN.md, and implements every screen as Storybook stories across the happy / empty / loading / error / AI states at production fidelity, plus an MDX flow-overview that sequences the screens into journeys. A brand-driven sibling of macosify's macos-storyboard skill (which is locked to a macOS-native look).
1 skill
Build a complete, reusable voice persona package for any person from examples of their real writing — the generalised factory behind the create-luke-content pattern. Mines a writing corpus into an evidence-anchored base voice (every rule traced to a quoted sample), layers purpose-tailored persona variants on top (LinkedIn posts as the flagship, grounded in bundled LinkedIn engagement deep research plus any user-supplied research; plus long-form, marketing, short-form, chat, email, code review, speaker notes), generates a per-person deterministic voice lint, and bundles a distilled 'Signs of AI writing' field guide (from Wikipedia's WP:AISIGNS) so everything the persona writes avoids every known AI tell. Delivers a ready-to-install create-X-content plugin skill or a portable docs folder. v1.1 adds a stylometric layer grounded in voice-replication research: word-count corpus grading, syntactic-fingerprint extraction (sentence rhythm, contractions, nominalizations) enforced as lint advisories, anchor-diversity and style/content-entanglement rules, NNG tone dials per variant, and long-form drift re-anchoring. v1.3 adds an owner-feedback maintenance loop (corrections as highest-grade evidence, fix-the-class discipline, retired-phrase mechanism) plus guards distilled from real owner reviews: signature-phrase rations, distribution-not-formula rhythm, referent discipline, audience-knowledge floor, register fencing, and a self-narrating meta-label tell. v1.4 adds the empirical research layer to the bundled AI-signs field guide (quantified structural tells from 2023-2026 corpus studies, the 99.6% expert-detection and AI Penalty evidence, structure-vs-texture discipline, epigram budgets, epistemic stance markers, flatten-and-reconstruct guidance) and threads the new guards through the package blueprint and self-check. v1.5 adds the feedback-disposition discipline (every review item applied, overruled with stated reason, or routed AND landed in the owner's ledger), the restatem
1 skill
Generate a single printable A4-page executive business case in Diolog's house style from a customer's proposal deck. Uses the /design-craft skill against a bundled single-page DESIGN.md and a templated version of the Alfabs business case, with all copy in the Diolog brand voice (create-diolog-business-case). Produces one dense, print-clean page (problem, solution, value, quantified outcome, investment, next step) on a fixed 794x1123 A4 canvas that prints to exactly one page.
1 skill
Build a branded, print-ready customer proposal / pitch deck from Diolog's 9-slide template system. Fills company-agnostic templates with a specific client's identity, contacts, pricing and numbers (prompting with multiple-choice questions for missing details), keeps the IR-proposal spine, and can design new slides into the same system via design-craft. Outputs a single self-contained web page with pan/zoom (trackpad + controls, fit-deck / fit-slide / 100%) that prints to a vertical, paged PDF at native 1920×1080 slide size.
1 skill
Self-contained slide-deck skill spanning three output targets from one body of deck craft: self-contained fixed-stage HTML presentations, lecturn.deck/1 JSON (and .pptx through the deckconv round-trip), and Diolog investor decks assembled from the 140-template / 12-recipe slide library. Carries its own visual-design layer rather than depending on a general design skill: consuming a supplied DESIGN.md or token file as binding context, authoring a direction when no brand exists (scheme / formality / density, subject-mined palettes, the signature element, the swap test), typography sized for projection with tracking and measure rules, colour strategy and accent budgets, five-vector hierarchy and cross-slide rhythm, anti-slop rules, and an accessibility floor calibrated for projection. Deck-specific narrative discipline: the title sequence as the deck's argument, speaking-vs-reading density, one idea per slide, deliver-the-whole-count, and the linear-flow usability rules (per-slide trunk test, recognition over recall, the honesty gate on charts and claims). Ships a per-slide build gate and a delivery review built on the verification epistemics that rendering an image is not seeing one, that the question you bring to a capture determines what you see in it, and that a clean gate means no known defect rather than verified. Direction evidence is read progressively from the template libraries on the machine (frontend-slides bold pack, open-design deck templates, portable design systems) without bulk-reading them. v1.1 folds the slide-template library into the skill: 200 layouts across 27 families, 21 recipes, the full slot contract and every validator rule. v1.2 closes the pipeline-integration gaps with the mandatory render_deck loop. v1.3.0 folds in portable ideas from impeccable. v1.7.0 turns the prose gate into an executable preflight probe. v1.8.0 hardens initial first-pass generation to eliminate design-review failure modes: establishes pure deterministic SVG charts by
1 skill
Turns a vague research need into a single copy-paste-ready Google Gemini 3.1 Pro Deep Research prompt — with pseudo-XML scaffolding, archetype-specific overrides (technical / competitive / regulatory / academic / forecasting), epistemic bounding tags, inline citation protocol, and wrap-around Operator Notes (Plan Review, decomposition, adversarial audit) that empirically matter as much as the prompt itself.
1 skill
Turn Claude into an opinionated, accessibility-aware, AI-slop-resistant design collaborator that produces intentional design artifacts in HTML/CSS/SVG/JS — landing pages, interactive prototypes, mobile app screens, slide decks, wireframes, hi-fi variations, design-token extractions, and component inventories. The SKILL.md carries a full design philosophy (content discipline, purposeful aesthetics that reject AI-template tropes, visual hierarchy and rhythm, typography and color systems, interface copy as design material, WCAG accessibility, interaction states, system thinking, respecting the medium) and routes to 29 phased procedures across production (discovery-questions, frontend-aesthetic-direction, wireframe, make-a-deck, make-a-prototype, make-tweakable, generate-variations, make-an-animation, make-a-doc, generate-images), craft (motion-design, gsap-motion, depth-and-3d, laws-of-composition, typesetting, data-viz, mobile-design), system (design-system-extract, component-extract, design-system-author), and review (unit-critique-gate, accessibility-audit, ai-slop-check, hierarchy-rhythm-review, interaction-states-pass, visual-verification, polish-pass). Built for Claude Code: AskUserQuestion for kickoff rounds, the Agent tool for ship-time review panels over real deliverables, and self-contained deck-scaling / device-frame / tweak-panel code that runs as standalone files. v1.2 adds mobile-design (iOS/Material platform grammar, thumb zone, input methods, industry conventions, emotional design), gsap-motion (timelines, ScrollTrigger scroll storytelling, SplitText, SVG draw/morph, Flip/Draggable — all plugins free via CDN), visual-verification (viewport-matrix layout integrity + screenshot playbook, a fifth polish-pass reviewer), an interface-copy chapter, subject-derived aesthetics with a plan-time signature element and swap-test genericness gate, and detection for all three current AI-default looks (warm-editorial, dark+acid-accent, broadsheet). v1.3 wires in local
1 skill
Reverse-engineer a product's visual design system from one or more website/app screenshots and produce a single authoritative DESIGN.md — a framework-agnostic, token-based spec (palette, typography, spacing, shape, motion, component inventory, voice, do/don't) with per-token provenance marks, so another AI (Stitch, Claude, v0, Cursor, AI Elements) can generate new screens that unmistakably belong to the same product.
1 skill
Generate a spec-compliant DESIGN.md for any live website from its real pages — the browser-driven, measure-first successor to design-md-from-screenshots. Given a URL it uses Obscura (parallel probe agents with rate-limit backoff, full-page screenshots over CDP + getComputedStyle-by-role measured tokens + a colour census, read from the spacing longhands) to MEASURE the design of the key pages (home, about, product/services, contact, and any investor hub/portal) instead of guessing hexes and fonts from a screenshot, then synthesises an authoritative DESIGN.md in the official google-labs design.md format (YAML token frontmatter + Overview/Colors/Typography/Layout/Elevation/Shapes/Components/Do's & Don'ts + per-token provenance + self-critique) and refines it into an intentional, upgraded identity.
1 skill
Build or extend a company website or standalone investor centre by composing gate-tested HTML blocks, theming them from DESIGN.md, filling copy from VOICE.md, and wiring Diolog live-data markers. Use for a marketing site, investor portal, site artifact, site structure + design + voice conversion, or when adding pages, sections, images, and explicit custom workspace blocks. Prefer this over hand-writing whole pages because the shared chrome, responsive system, and stock blocks already pass the Studio gates. Triggers include "build the company site", "generate the investor centre", "scaffold a marketing site", "make a site from this DESIGN.md/VOICE.md", "add a pricing page", and "theme this template to the brand".
1 skill
Analyze product discovery and customer feedback session documents using the Discovery Sentinel persona. Types sessions by purpose (discovery vs demo), segments distinct client companies, and produces lean Feedback Classification Reports, Discovery Insight Briefs, readable Prioritised Opportunity Assessments, an Action Checklist, and company-specific Sales Discovery Companions — with confidence scoring, JTBD extraction, bias detection, and compliance flagging for regulated B2B SaaS. Includes a trend-synthesis mode across prior analyses.
1 skill
Compress a verbose source document — README, API docs, research paper, technical guide, library docs — into a maximally dense, semantically complete Markdown reference (written to a `<source>.summary.md` file) that another LLM or a skimming human can consume in a fraction of the tokens. Uses a triage-tag system ([CRITICAL]/[WORKFLOW]/[POWER-USER]/[GOLDEN-NUGGET]/[DEPRECATED]), contextually integrated tips, solution-first problem framing, multi-tier feature capture, and a mandatory Strategic Synthesis footer (Maturity Model + Feature Selection Matrix + Workflow Progression Ladder). Target compression 60–80% while retaining 100% of semantically unique information.
1 skill
Turn a named Diolog feature (or several) into email-ready product-mockup graphics on a Figma canvas. Reads the feature's real context — docs/marketing/product-feature-guide.md + features-build/plain/NN-*.md + the rendered mock UI read in a browser via Obscura (the served web / customer / investor design-system hosts) — then designs and builds, via the design-craft skill, impression-not-replica mockups that lead with the feature's payoff (a full product-window app shot — the complete app shell in a clean window — for hero features, plus a product-surface vignette, soft panel + peeking device, collage/overlap, phone bezel, and a website frame for actual public sites) and make the product look genuinely smart (a real insight, an italic quote from the record, named source chips) as titled, spaced artboards in one self-contained HTML file using the live design tokens, and recreates that layout in Figma via whatever write-capable Figma MCP is connected. Each artboard also carries a brief non-technical marketing caption beneath it for IR teams and the agencies who serve them (no em-dashes), with an optional email-hero split format, and semantic emphasis via a soft wash and icon rather than AI-slop left-only borders. If no Figma MCP is present, the HTML file is the final deliverable.
1 skill
Turn messy, multi-source customer-discovery/feedback dumps (Slack exports, Google-Drive "Notes by Gemini" .docx meeting notes, segment-foldered research/ICPs/personas, AI-Studio JSON repos, screenshot evidence, audio/video) into ONE clean, deduplicated, verbatim Markdown archive organised by buyer segment with an INDEX — the pre-step that feeds discovery-sentinel and product-strategist. Converts docx/pdf/txt via pandoc, classifies meetings by attendee email domain + source folder, dedupes identical content + notes/transcript pairs + Slack recaps, mines Slack for product feedback (dropping bots/internal-eng/test noise), strips Google/Gemini auto-summaries only where a transcript exists, and builds a who's-who INDEX with provenance and probable dates.
1 skill
Generate a complete investor-portal record — theme tokens, content, imagery and motion — into the Diolog database from a company's DESIGN.md and company overview.
1 skill
Review and rewrite prompts, system prompts, agent instructions, CLAUDE.md files, and Claude Code skills against Anthropic's prompt-engineering guidance and the Karpathy coding-behaviour principles. Diagnoses the artifact with a structured rubric (clarity, examples, XML structure, output format, hallucination guards, consistency, guardrails, token weight, rule craft, and — for skills — trigger description quality and progressive disclosure), then delivers a surgical rewrite with per-change rationale. Bundles the full Anthropic prompting docs (best practices, Opus 4.8 guide, guardrail guides, four production use-case guides) and the Karpathy material three ways: the guidelines verbatim, worked before/after examples, and a prompt-craft dissection (karpathy-prompt-craft.md) that treats the guidelines as a model of rule-writing — principle anatomy (name, bold directive, checkable behaviour bullets, self-test), tradeoff clauses, imperative-to-verifiable transforms, ownership-scoped cleanup, working-if success signals, a cross-surface reuse pattern (one canonical rule text wrapped per surface), and a mergeable hardening block — which the skill applies to every prompt it improves.
1 skill
Create Diolog Tasks issues from a code-review markdown report. Each `### [SEVERITY] Title` finding becomes one issue with severity (Critical/High/Medium/Low) and type (Vulnerability/Bug) labels and Todo status. Formerly log-bugs-to-linear; now targets the diolog-tasks MCP. Preserves the reviewer's exact prose, file:line citations, fix recommendations, and confidence scores.
1 skill
Incrementally digest macOS app UI screenshots, app icons, and official Apple UI kits (user-supplied, e.g. from macapp.supply; .sketch kits are deconstructed directly from their JSON) into a persistent design corpus — per-app token profiles with framework-lineage classification (AppKit-native vs Catalyst vs iOS-on-Mac vs web-Electron; only native evidence feeds canon), cross-app pattern entries, aesthetic style clusters, icon anatomy digests, and a master TASTE.md — powered by a Mac Design Archivist persona, a deep-research knowledge base (14-point screenshot rubric, anti-pattern taxonomy, WCAG/HIG thresholds), and a macOS-native grammar reference (Liquid Glass layer discipline, selection/accent/density tells, concentric-corner geometry, 10-point native-tells audit). Then designs beautiful, native-feeling macOS mocks and icons from the corpus, auditing its own output against both rubrics before delivery. Recalibrated for current-model behaviour (Anthropic's Claude Opus 5 prompting guidance): design mode's pre-delivery audit now opens the render before scoring it (a rubric scored off the source is scored off the wrong artifact — render, crop each region at DPR 2-3, read the crops asking 'what is wrong with this?', first-person rather than delegated), and the batch summary's deltas-only rule is stated as a length rule too, since written output drifts long by default.
1 skill
Design beautiful, varied, authentically native macOS app UIs and app icons from scratch — full mock windows (HTML/CSS or token-precise spec) and 1024pt icon designs (SVG or Icon Composer layer spec), each committed to a named aesthetic direction drawn from a catalogue synthesised from 135 curated macapp.supply apps and 134 icons, grounded in Apple's macOS 27 UI kit (control ladder, type ramp, label tiers, chrome metrics) and the HIG. Routes native correctness to the macosify HIG library and craft/variation/polish to design-craft and ux-craft, audits every delivery against a 10-point native-tells audit and 14-point quality rubric, and enforces variety discipline so consecutive designs don't converge to one template. Every design declares one signature element and one justified aesthetic risk (with an AI-default calibration guard on the looks generators overuse), directions are calibration not a closed menu (hybrids and novel directions welcome when the subject earns them), and an incorporated fluid-interface motion layer (springs, interruptibility, materials-on-web, optical typography, reduced-motion floor — from Apple's WWDC design corpus) covers interactive mocks and ships a motion spec with static ones. v1.2 adds the essence layer: references/mac-essence.md distils the eight convictions of great mac app design from the corpus + UX canon (familiarity-as-canvas, restraint-as-confidence, the surface's single question, states-are-the-quality, words-as-load-bearing, forgiveness-over-interrogation, the keyboard, invisible completeness) plus a yield table where design-craft/ux-craft web-first rules bend to mac grammar (cursor, targets, casing, toasts); the build step now designs states-not-screens (ideal + empty rendered, full state matrix specified with real unhappy-path copy), writes the real words per macOS copy grammar, names keyboard behaviour, and delivery adds the essence test (name the question, the signature, and the worst state's behaviour); ux-craft is elevate
1 skill
Keep the Diolog marketing/feature documentation set in the diolog-team-files repo (~/Dev/diolog-team-files/web) current when a feature ships or a Diolog Tasks ticket lands. Updates the two assembled published docs — existing-features-technical.md (full technical detail) and exhisting-features-marketing.md (plain, non-technical) — plus the per-area detail under existing-features-references/ (final/ + plain/ numbered per area 01-22, backed by the live/ and raw/ capture layers), and the auxiliary surfaces outbound-contact-surfaces.md, ui-interaction-details.md, and INCORPORATED.md (reconciliation ledger). Enforces the content standards (technical files carry component names, routes, GraphQL ops and exact copy with no opinions; plain files carry zero technical terms, second person, sentence case, no em dashes, no emojis), pair consistency (each assembled doc agrees with its per-area source), the supersede-don't-accumulate currency rule, and the live-app > source > ticket source-of-truth hierarchy. Operationalises docs/marketing/MAINTENANCE.md (the canonical standards guide in the app repo), with the published docs now living in diolog-team-files/web.
1 skill
Validate that an implemented React or React Native UI faithfully reproduces a reference mockup (HTML+CSS, served prototype, Figma export, or another rendered component set), then update the code to close every gap — by measuring the rendered tree, never by eyeballing or reading source. Treats the mock as the source of truth, builds a COMPLETE screen/frame inventory (never silently drops a frame), extracts both sides to on-disk artifacts and generates the ledger from them, inverts the burden of proof (a visible difference is a defect until a citation proves it intentional), diffs structure-first then per-property computed styles, and produces a per-screen present/divergent/absent ledger plus a FUNCTIONAL-GAPS document for any newly-added UI that is only visual (missing backend, data wiring, endpoints). For React Native it measures the RENDERED tree — structural/accessibility extraction (axe describe-ui / Maestro hierarchy) plus resolved style props over the Metro CDP connection (rozenite / agent-cdp) — and forces every screen's measurement to disk as artifacts, so a 'match' verdict can never come from a screenshot, a source read, or reasoning (frontier vision models catch under half of fine-grained UI differences; an artifact-forcing gate plus a completeness-critic close that hole). For a React web target the comparison is symmetric DOM-to-DOM — reference and target are both read via getComputedStyle through one extractor at the same viewport, with the app chrome (sidebar/header/nav) in scope. Asks the user when unsure, on both a full-comparison and a per-screen basis. Generic and framework-agnostic — nothing app-specific. v2.6.0 re-platforms the capture harness onto Obscura over CDP (odiff-bin is now its only dependency): the element-scoped raster and IoU text-less layers stand, and the CDP rendered-font layer is recorded as available:false with its reason rather than as zero divergences, because Obscura loads no web fonts and every named family collapses onto one f
1 skill
End-to-end product-positioning pipeline. From a company context it generates two Google Gemini Deep Research prompts (a positioning brief + a product-research-persona market brief), renders a copy-and-launch HTML with animated Gemini instructions (select Deep research from the + menu; Share & Export → Copy contents), then — once the research comes back — synthesises three framework-grounded positioning territories (Ries & Trout, April Dunford, Blue Ocean, David C. Baker) as Markdown plus a stylised interactive decision page.
1 skill
Reverse-engineer a slide deck's visual design system from one or more slide screenshots and produce a single authoritative PRESENTATION_DESIGN.md — a tool-agnostic, token-based spec (canvas, palette, typography, slide archetypes, chart style, image treatment, iconography, motion, do/don't) hard-capped at 1,700 characters so another AI (Gamma, Beautiful.ai, Pitch, Keynote/PowerPoint co-pilots, Stitch, v0, Claude, Cursor) can generate new slides that unmistakably belong to the same deck. Sibling to design-md-from-screenshots but tuned for slides: fixed 16:9 canvas, slide archetypes, chart data-ink discipline, image treatment conventions, and on-slide positional grammar.
1 skill
Whole-product gap analysis for any project against any stated goal (v1.0, GA, a roadmap horizon, or the product's own stated intent) — 'where does this project stand and what's left?'. Surveys everything the repo has — status ledgers/orchestrator logs OR issue-tracker state, specs' deferral sections, untriaged briefs or roadmap docs, reference catalogues, the live app code, navigation/wiring, and the test inventory — via up to 7 parallel research agents (status-map, catalogue-currency, deferral-mining, backlog-classification, wiring-audit, test-coverage, product-coherence), adapting the lens plan to whatever artifact classes exist. Cross-checks every status claim against git (ledgers lie; git is authoritative), reconciles contradictions with a fixed authority order, and synthesizes one comprehensive evidence-cited gap report stratified into seven layers (coherence gaps ranked by contradiction-with-the-product's-own-claims, genuinely-unbuilt features, deferred tails + shared substrates, security/legal release gates + dark-by-default workers, test debt, docs drift, live defects) ending in a prioritized P0/P1/P2 proposed-work-items table that seeds the repo's pipeline (briefs or tickets). Read-only: writes one report, never product code. Built for reliable execution by Opus-class models — explicit agent prompt templates, epistemic laws, and a fixed report skeleton.
1 skill
Generate validated, prioritized product feature concepts and outcome-based roadmaps for AI-assisted IR/compliance/governance SaaS. Supports feature ideation, roadmap construction, feature evaluation, data ingestion, and comparative prioritization using RICE-C/WSJF scoring, compliance gates, and OST methodology.
1 skill
Recursive deep-research pipeline with two tracks. Track R (product roadmap): grounds in the product's existing surface, authors strategic + high-volume-ideation + ~10 drill-down Gemini Deep Research prompts (exclusion blocks, analysis lenses, epistemic bounding), then synthesises the returned reports — via verdict ledgers, an evidence hierarchy, gates-track sequencing, and red-team discipline — into wave-based roadmaps with stable numbered, triage-ready feature entries. Track G (general): takes ANY topic, thought, or idea → one initial research prompt (via deep-research-prompt-creator) → pause for the user's Gemini run → 5–10 drill-down prompts from the returned report → pause → synthesis into a decision-ready document.
1 skill
End-to-end feature delivery conductor for your product. Takes ONE feature (inline description or a markdown file) and drives it from a bare idea to merged, tested, production code by orchestrating the existing skills in a single in-session flow: design-craft (represent the whole feature UI in the design-system elements/composites/mock-pages, grounded in the repo's design system — e.g. apps/web-design-system — plus an optional DESIGN.md design-language spec when provided or present, and any provided HTML/CSS/JS mock) → triage (→ docs/specs/spec-DIO-XXXX.md) → plan (→ docs/plans/plan-DIO-XXXX.md) → work (implement in an isolated worktree) → a deferred-work loop (re-work, or child triage+plan+work for extra phases, all on the SAME branch) → gap-fix (close remaining spec gaps in code) → acceptance-e2e (a comprehensive e2e suite over every user flow/action/interaction/menu, run against the feature branch locally and fixed) → finalize (commit → rebase → merge into the integration branch → push → clean up worktree(s)). The conductor stays in-session so it keeps the accumulated feature context across every stage and leans on the on-disk spec/plan/mock artifacts as durable, resumable memory; parallelism lives inside each sub-skill's own Workflow fan-out. One feature = one branch = one worktree; the merge is behind a fail-closed pre-merge gate. Builds on the design-craft, feature-spec-pipeline (triage/plan/work/gap-fix), and acceptance-e2e plugins. v1.5 recalibrates the conductor for current-model behaviour (Anthropic's Claude Opus 5 prompting guidance): 'Verify, don't assume' becomes verify-with-oracles-not-by-re-reading (gates actually run, real paths exercised un-stubbed, out-of-family gates — but no separate pass whose only job is re-reading a stage's own output, which now costs tokens without adding recall), plus a front-load-the-whole-brief discipline (a complete spec handed over once and left to run is the condition the pipeline performs best under, which is what makes
1 skill
Backlog-wide feature-delivery orchestrator. Surveys ALL remaining feature work in a product repo — LEDGER.md, docs/specs/spec-*.md, docs/plans/plan-*.md, untriaged briefs in docs/features-to-triage/, refined HTML mocks in design/mocks/html, deep research in docs/deep-research/ — reconciles any in-progress worktrees, then writes ORCHESTRATOR.md (a combined plan + ledger for everything left) and orchestrator-hierarchy.html (a visual dependency hierarchy) BEFORE any execution. It then drives the backlog to done by running the ship-feature skill per feature: dependency-ordered, up to 8 concurrent Opus runner agents, with mechanical plan-scoped coding optionally delegated to the Cursor CLI (composer-2.5) and verified/fixed by Opus, while the orchestrating session itself stays on the session model. Includes a preflight that checks the project's directory/file conventions (offering to clone diolog-team-files for CODING_PRACTICES.md / NEW_PROJECT_BEST_PRACTICES.md, gather stray feature briefs into docs/features-to-triage/, and validate the monolith layout against NEW_PROJECT_BEST_PRACTICES.md). Builds on the ship-feature plugin (and through it design-craft, feature-spec-pipeline, acceptance-e2e). Recalibrated for current-model behaviour (Anthropic's Claude Opus 5 prompting guidance + the effort docs): the runner wire self-check no longer hardcodes a dated model id (a check pinned to last generation's string fires WRONG-MODEL on every correctly-routed runner and stops the fleet before it starts — it now names the TIER, so a newer Opus passes and a wrong tier fails); launch rule 2 gains the effort dial (an agent spawned with no effort runs at `high`, `low` is the level built for subagents, a feature runner is the long-horizon shape `xhigh` exists for — pair it with `max_tokens` ≥64k — and a review lane should drop effort before model, since REVIEWER ≥ WRITER is about capability class); the warm/cold resume boundary is corrected from a ~5-minute prompt-cache TTL to the harness
1 skill
Validate whether a specification, build plan, or Definition of Done is GENUINELY implemented in a large codebase — not just rendered, schema-valid, and passing tests. Classifies every claimed-done data field as REAL (a producer computes it from the filesystem / git / discovery / AI / a live API), AUTHORED (read from persisted state but only ever written by a UI round-trip or a seed — nothing computes it), or MOCK (only ever a fallback constant), each with file:line evidence. Fans the work out across parallel investigators, adversarially verifies the highest-impact findings before reporting, and writes a ranked, date-stamped living gap report that doubles as the resolution backlog. Built to catch demo-data-dressed-as-done and to re-validate after a prior '100% complete' claim — the core insight being that renders + schema-valid + tests-pass never proves a real producer exists. Audit-only: it changes no product code — the gap-fix skill (feature-spec-pipeline) is the fixer.
1 skill
Session-start operating defaults that cut token spend without cutting quality, tuned for Claude Opus 5. Targets where the tokens actually are — 96% of a session's bill is reused input, resent from the top every turn — so it addresses repetition, delegation and read width rather than prose length. Ships as a ~1029-byte literal designed to sit inside the cached system prefix, paid for once per cache window rather than re-read. Includes the reasoning for every rule, the four alternatives that were rejected as traps (tool-definition stripping, transport-layer prefix truncation, request-hash response caching, and a plain 'be concise' instruction), an explicit measured-vs-assumed split, and a comparison against prose-compression skills like caveman with the four-arm benchmark that would settle it. Perch injects this at session start behind a toggle; it also works standalone.
1 skill
Book-grounded UX engine for implementing, mocking, and reviewing web and mobile UIs, layouts, and user flows, plus marketing and transactional emails. Operationalizes the UX canon — Krug (self-evidence, trunk test), Norman (affordances/signifiers/feedback), Garrett's five planes, Cooper's goal-directed design, Nielsen's heuristics, Wroblewski's mobile-first, Yablonski's Laws of UX, Weinschenk's behavioral science, Fogg/Cialdini/Eyal persuasion with a hard ethics gate, Tidwell patterns, Lidwell principles, Refactoring UI hierarchy tactics, Lean UX mock-to-learn — as three modes (Build / Review / Advise) over eight references: psychology-laws (research-cited mechanisms), review-playbook (multi-lens pass with severity honesty, intentional-vs-accidental deviation handling, evidence-format findings with pasteable fixes, and prompt-injection guards for reviewed content), flows-and-forms (GDS one-thing-per-page, validation timing, state machines, undo vs confirmation, wayfinding), mobile-ux (touch targets, thumb zones, iOS/Material platform grammar, RN notes), email-ux (two-second triage contract, client rendering constraints, transactional patterns, CAN-SPAM/GDPR compliance floor), ux-writing (microcopy patterns, tone matrix, Butterick typography), ai-product-ux (AI inputs/wayfinding/governors/narration/trust/tuners with friction-to-risk calibration and prompt-injection-aware retrieval rules), and per-surface pre-ship checklists incl. WCAG 2.2 quick pass and an AI-feature gate. The UX brain to design-craft's visual hands: it owns whether the thing works for humans; design-craft owns whether it looks intentional. v1.1.1 adds two pre-ship checklist floors routinely forgotten in mocks: designed not-found/error routes (no dead ends) and legally required links (privacy/terms/cookie consent). v1.2 adds a Working-posture section recalibrated for current-model behaviour (Anthropic's Claude Opus 5 prompting guidance): reviews run as two passes that must never be merged — find wide
1 skill
Find recent dynamic workflow runs that lost agents to API errors, and resume or finish them correctly. A workflow agent that hits a rate limit, usage limit, dropped connection or 5xx returns null with zero retries; parallel()/pipeline() map that to null too, the generated script's closing .filter(Boolean) drops it, and the run reports `completed` with the loss unreported. This skill separates the two shapes that need opposite handling: a finished run with dead agents, and a run that is still live (which must not be relaunched, and whose worktrees and pid-held locks must not be touched). Bundles a stdlib-only read-only scanner that reads the journal and per-agent transcripts rather than the snapshot, because a snapshot is written once at completion and is absent or stale for any live or resumed run; it reports run id, session id and whether that session is still alive, the scriptPath and args needed to resume, journal started/results counts, and each agent's item, state and error. Encodes the mechanics verified from the claude 2.1.220 bundle: the journal is filed under <SESSION-UUID>/subagents/workflows/<runId>/ so a changed session id cold-starts a resume silently (anthropics/claude-code#65796, open) and the fix is a cp -R relocation; the cache key is a sha256 chain over (previous key, prompt, normalised opts) with a sticky miss flag, so replay is a prefix and a run that failed early recovers almost nothing; and the /workflows `r` key only aborts a live AbortController, so a failed agent cannot be retried there. Also carries the reconciliation discipline that matters most in practice: cached results are what an agent said, not what landed, and runs have reported items merged whose branches were still tens of commits ahead of main, so branch ancestry and worktree state decide whether a failed item is resumed in place or re-run clean. v1.1.0 corrects three defects the 2026-08-21 Ghostty crash exposed, each of which made a recoverable run look unrecoverable: liveness n
1 skill