jdh-agents

Curates, catalogs, retrieves, and maintains your library of notes and knowledge in an Obsidian vault. Workflow-shaped (Capture / Process / Retrieve / Maintain) and agent-mediated (skills draft, agents execute). 19 skills + 4 agents + 6 references, coupled to the vault conventions in ~/Loose Ends/.claude/rules/{wiki.md,catalog.md} and the Templater templates in ~/Loose Ends/Templates/.

19 skills · 4 agents

Software craftsmanship discipline skills: grill (the interview primitive — walk a design tree one question at a time, facts looked up by subagent rather than asked, no acting until shared understanding), grill-with-docs (that loop applied to vocabulary, against per-repo CONTEXT.md glossaries), diagnose, prototype, tdd, improve-codebase-architecture, grok (stateful multi-session codebase comprehension with supersession-aware understanding records), zoom-out (adapted in part from mattpocock/skills, MIT), infra-review (architecture/security review of Terraform+Atlantis PRs), quiz-me (opt-in active-recall comprehension check after a code explanation), interrogate-model (holistic representability review of a domain model — the legitimate scenarios and personas it cannot express), design-by-stories (design an artifact's shape — schema, template, data model, config — by modeling its actors and user stories first, then deriving the shape from the stories), and lucid-architecture-diagram (drive AWS cloud architecture diagrams in Lucid, including the two silent-failure modes of its API). derive-conventions (derive a repo's unwritten house style from its own code, adjudicate each inferred rule with the human, and write the approved set to CONVENTIONS.md so the review agents stop re-deriving it). wizard (author an interactive bash wizard that walks a human through the steps only they can perform — third-party dashboards, credentials, CI secrets, one-off cutovers — from a hardened template), resolve-conflicts (recover each side's intent from commits, PRs, tickets and NDR heads before resolving a merge, rebase, or jj conflict; detects git vs jj and follows that VCS's own mechanics), and wait-what (user-invoked: re-pitch a message that did not land, in Simplified Technical English using the repo's own vocabulary) — all three adapted from mattpocock/skills, MIT. screen-critique (critique a UI screen design in a fixed order — purpose against its neighbouring screens, then structure,

19 skills · 3 agents

ADHD-friendly productivity coaching: daily planning (/today, /energy), end-of-day capture (/checkin), weekly scheduling (/plan-week) and review (/weekly), life-direction reviews (/review), goal-to-project alignment (/align), project lifecycle (/intake, /breakdown, /reentry, /sunset), idea capture (/spark), brain dump (/dump), task triage (/triage), decision journaling (/decide), and pattern analysis (momentum, overdue-rescue) with adaptive tone calibration

16 skills · 3 agents

Personal SDLC scaffolding for AI-assisted development. A lightweight, ephemeral 'contract' between user and AI drives each code change from captured idea to verified done; the durable layer (CLAUDE.md, ndr atoms, README) lives elsewhere. Lifecycle: capture -> draft -> implement -> close.

7 skills · 2 agents

PM helpers for team Linear workflows. Skills for a PM-mode planning session (read the real state, slice work into cold-handoff tickets, confirm every write), single-ticket authoring (a short shaping interview, body composed to the shared issue-body spec), the weekly grooming loop (per-person WIP), plan-to-tickets breakdown (assignee-aware), chart-first mapping of efforts too large to plan in one session, questionnaires that pull a blocking decision out of someone else's head, and end-of-cycle retro notes (attribution + Linear-first output). Composes with the linear plugin; ndr and librarian integrations are optional externals.

7 skills

Interpret Klippain Shake Tune results for 3D printers: diagnose issues from PNG graphs, explain findings, recommend fixes, and track results over time. Supports belt comparison, input shaper, vibration profile, axes map, and excitation tests.

1 skill · 5 agents

A suite of conversational thinking tools — reflect (Socratic mirror, never recommends, always saves), mull (thinking partner, probes AND contributes opinions, saves substantive sessions only), and converge (advisor, leads with a researched recommendation plus confidence % and refines it one question at a time). Plus a paired evaluation track: first-principles (evidence-first interview that produces a signed-off needs map before any option is named) hands off to solution-research (turns that map into a systems-first parallel-research handoff prompt).

5 skills

Dialectical decision analysis: dispatch parallel advocate agents to research opposing perspectives, optionally fact-check claims, challenge the consensus with devil's advocacy, and independently synthesize a verdict. Three modes (Quick/Standard/Deep) control pipeline depth. Not for factual lookups or simple preferences.

1 skill · 4 agents

A two-sided loop for improving agent plugins from test sessions. The 'session' skill (tester side) turns the current session record into a copy-pasteable report grading the plugins exercised; the 'jot' skill files a quick note or idea about a single surface mid-session; the 'triage' skill (author side) aggregates many reports and jots into a prioritized action plan, routing each finding to a concrete fix. All three share one report-format contract.

3 skills

Linear workflow conventions for a small two-person team. Owns ticket creation defaults (including assignee), status flow semantics, label set, title shape, collaboration conventions, and MCP call patterns. Composes with spec-flow (which uses Linear as a contract host but defers to this plugin for the conventions). Includes a doctor skill for workspace setup validation, and a linear-ops agent that executes decided ticket writes in isolated context — resolving identifiers, applying the MCP silent-failure workarounds, and verifying what landed.

2 skills · 1 agent

Craft and upkeep of skills. Authoring reference (the vocabulary and principles that make any agent-consumed document predictable — a skill, an AGENTS.md/CLAUDE.md, a rules file, a doc reached by a pointer) plus provenance tracking for borrowed work: compares an adapted skill to its pinned upstream, flags behavioral drift and fabricated attributions, and refreshes the reviewed commit SHA. Houses writing-for-agents and upstream-review.

2 skills · 1 agent

Study foreign open-source projects to extract transferable architectural lessons into the Obsidian vault. `/teardown` runs a study session over a local clone, or over files fetched at a pinned SHA when no clone exists, dispatching a read-only `surveyor` agent per project through two ordered phases (current structure, then history/rationale) descending one user-picked layer at a time (system map, domain model, code), spot-checking citations before any write, and adding each layer to an interactive Artifact page and to one of two note types: a single-concern QUESTION note that accretes comparison rows across projects over time, or a whole-system PROJECT note anchored on a mental model. Both carry `projects_surveyed:` frontmatter pinning the repo, commit SHA, and date read — the index and the read log in one place. `/teardown-recheck` audits that pin against the corpus's current HEAD, diffing cited files since the recorded SHA and verifying that cited line numbers still point at what the note claims, so drift is caught instead of trusted. A reference instrument for lookup, not a learning exercise. Composes with librarian (graduation, refresh) and ndr (decisions about the user's own projects); distinct from craft:grok, which studies repos the user owns and intends to change.

2 skills · 1 agent

Format-aware atomic commits for git and jj (Jujutsu). One skill covering commit, split, message review, and retrofitting edits into earlier commits (jj absorb / git fixup). Auto-detects VCS and commit message style from repo context. Includes PreToolUse hook that blocks destructive working-copy operations.

1 skill · hooks

Engineering-manager skills for executing tracked work, with no fixed lifecycle. `scope` scopes one ticket (fetch, ground in decisions, read the touched code, ask only the questions nothing else can answer) and stops at a brief. `lead` takes any work reference, checks from the ticket body whether it is already shaped, picks an execution shape (inline, one worker, parallel workspaces, test-first), dispatches, and verifies every result by diff before a pre-push stop. Tracker-agnostic: Linear, Fibery, and GitHub Issues through per-tracker adapters, selected by the reference form or a repo-local `.em.toml`. Pairs with the pm plugin, which shapes the tickets em executes.

2 skills

Introspection tooling for Claude Code itself. usage-report: reconstruct how a person works with Claude Code in a repo (skills, slash commands, subagents, MCP tools, task-kickoff style, activity over time) from local Claude Code transcripts, via a stdlib-only read-only parser. It can be invoked from Claude Code or Codex, but does not parse Codex rollout files or claim Codex usage parity. conversation-temperature: read the local user's own authored register across all their Claude Code transcripts — terseness, heat, and warmth, led by marker rates and real quoted messages rather than composite indices. Both run locally against the user's own transcripts and make no network calls; usage-report emits names and counts only, while conversation-temperature quotes the user's own messages back to them by design.

2 skills

Drive Penpot through its MCP plugin without the silent failures: text edits that render stale until forced to re-layout, async page switching, suspended-tab hangs, and saving named versions before irreversible edits and after settled design decisions.

1 skill

A stateful, multi-session teaching workspace routed into the Obsidian vault. The /teach skill grounds learning in the user's documented learning style and high-trust sources, builds a mission, resources, glossary, and learning records as vault notes under the topic's best-fit context, and writes lessons as markdown notes that transclude the wiki's prose and add a pure-CSS self-grading quiz (via a shipped snippet, no plugin or JS). Cheat sheets and reference material live as ordinary wiki pages, not HTML. Entry altitude is set by question altitude and graph density, with learning records as the only gate on mastery claims. Grounds in owned textbooks via DEVONthink when available and pins live-code citations to immutable refs. Requires an Obsidian vault (obsidian-cli); optionally surfaces owned textbooks from DEVONthink. Adapted from mattpocock/skills (MIT, © 2026 Matt Pocock).

1 skill

Teaches agents that isolated per-agent checkouts exist in jj-colocated repos via `jjx` (jj's equivalent of a git worktree), and how to drive it: `jjx open <slug>` creates or idempotently resumes a managed sibling workspace and prints its path on stdout, `--from <revset>` bases it on an explicit revision instead of the caller's uncommitted state, `jjx open --unique <slug>` forces a fresh workspace for fan-out, `jjx add <path>` adds an existing checkout at an explicit path, and `jjx rm <path>` cleans up without discarding unsaved work. The single skill exists to be found by the failure it prevents — an agent about to conclude that parallel isolated work isn't possible in a jj repo, or reaching for a plain git worktree instead. A Claude Code mod shows which subagents work in which workspace: a pane grouped by workspace, opened by `/spaces` or when a subagent first lands in one, placing each agent from the `<repo>-spaces/<slug>` paths in its spawn prompt and its own tool calls.

1 skill

A Claude Code mod that watches the pull requests a session opens and the CI on their heads, on GitHub and Forgejo. A successful `gh pr create` or `fj pr create` adds the PR to a pane; a 20-second timer then asks the forge's own CLI for state and checks (`gh pr view --json statusCheckRollup` on GitHub; `fj pr view` plus `fj actions tasks` matched to the head commit on Forgejo), so each CLI's existing login is the only credential involved. A toast fires when CI passes or fails and when a PR is merged or closed. `/prs` opens the pane, `/prs add <url>` watches an existing PR, `/prs refresh` polls now, and `/prs clear` drops merged and closed ones.