workbench-core

Plugin workbench-core from marketplace workbench in HunterSpence/claude-workbench. Commands read from the repo's manifest — 3 days ago.

Skills (39)

• api-bridge

  Wrap an existing REST or OpenAPI API as a curated MCP server. For a server from scratch use mcp-builder.

• app-navigator

  Add an in-app assistant that answers \"where do I do X?\" with links into a signed-in web app.

• batch-review

  Reviews, judging or bulk passes via the Message Batches API at half price; needs an organisation API key.

• brain-audit

  Read-only health check of the Claude Code setup and knowledge base: stale archives, backup files, dead @imports, missing skill or hook targets, credential-looking settings, oversized always-loaded files. Use for \"audit the setup\", \"brain-audit\" or after a config change.

• brand-spec

  Ask one question at a time, then write BRAND_SPEC.md with locked design tokens.

• cascade

  Use for any non-trivial fix, build, refactor, migrate or harden task (fix or debug X, it is broken, add a feature): diagnose, plan, implement, verify across personas, with parallel-slice and best-of-N modes. Skip one-line fixes and pure research or decisions.

• cleanup

  Archive confirmed dead code into _archive/ with a restore manifest. Never hard-deletes; confirm-gated.

• ctx-agent-history-search

  Search local coding-agent history (Claude Code, Codex transcripts) with ctx before acting; ctx blame traces a file, line, commit or PR to its session. Use when prior decisions, attempts or failures may matter, or for a cross-session history report.

• deep-research

  Deep, verified web research ending in a decision list (critical / yes / maybe / no). Use for \"deep research\", \"research in depth\", or a question needing many sources checked (vendor or tool evaluation, \"should we adopt Y\"). Not for a single fact or code search.

• diagram

  Turn an English description or Mermaid text into a diagram: .mmd source, editable .excalidraw, SVG and PNG, rendered offline in a local browser. Use for flowcharts, sequence, state and architecture sketches for docs, decks and pull requests.

• draft-reply

  Draft a reply to an email or chat thread in the user's voice: answer first, then asks with owners and dates, as plain text and HTML or a .eml file. Never sends. Use for \"draft a reply\", \"help me respond\", \"write back to\", \"answer this email\", \"reply to this thread\", or a pasted thread, .em

• find

  Find files and facts across work libraries, captured pages and the how-to library: file-name catalog, keyword plus semantic search, related items, newest version. Use for \"find the doc about X\", \"what do our files say about Y\", \"latest version of Z\". Cites file#pN.

• github

  GitHub through the gh CLI, including Enterprise Server, GHE.com tenants and Managed Users: pull requests, issues, Actions runs and logs, reviews, releases, API queries. Use for any GitHub task or \"which gh command\" question. Not for local-only git.

• handoff

  Turn a mission into a self-contained handoff document a fresh Claude session can execute, or a short resume note. Use for \"write a handoff\", \"resume note\", a usage-limit warning, or work that must survive a restart.

• investigate

  Systematic debugging: no fix before a proven root cause. Use for \"debug this\", \"why is this broken\", \"investigate this error\", \"root cause\", a stack trace, a flaky test or unexpected behaviour. Hand a multi-file fix to cascade once the cause is known.

• kit-update

  Update the workbench kit, then verify it. Use for "check the new stuff and go", "update the kit", a carried release zip, a pending post-update verify, or approvals off. Not for updating Claude Code.

• linux-service-triage

  Diagnose a failing Linux service from evidence: systemd or PM2 status and logs, ports, disk and memory, Nginx proxy, DNS, TLS; propose a minimal fix with rollback. Use when a server app, container or site is down, erroring, slow or unreachable.

• mcp-builder

  Build and test an MCP server from scratch. To wrap an existing REST API use api-bridge.

• mcp-heal

  Probe every MCP server and report OK / DEGRADED / DOWN with the fix for each.

• office-docx

  Build a Word document (.docx) from an outline, Markdown or a corporate template. Use for \"write a memo\", \"make this a Word doc\", \"use our template\", \"turn notes into a report\".

• office-pptx

  Build or rebuild a PowerPoint deck (.pptx) with python-pptx from an outline or corporate template (tables, editable charts, images, speaker notes), then render every slide and fix it. Use for \"make a deck\", \"turn this into slides\", \"use our template\", \"add a chart slide\".

• office-render

  Render a .pptx, .docx, .xlsx or .pdf to one PNG per page for visual QA (\"does the deck look right\", \"render this document\"), and make a scrubbed copy of a file.

• office-xlsx

  Build or edit Excel workbooks (.xlsx) with openpyxl like a careful analyst: live formulas, input colour convention, named styles; recalculate and check for formula errors. Use for \"make a spreadsheet\", \"build a model\", \"add a summary sheet\", \"fix this workbook\".

• onboard

  Short interview that fills in USER.md (your work identity) and seeds the knowledge base.

• pc-checkup

  Read-only health check of this Windows PC with a ranked problem list. Run only when asked.

• pdf-report

  Build a polished PDF report, memo or handout from HTML or Markdown with Edge, check it, then merge, split, stamp or fill PDFs. Use for \"make a PDF report\", \"merge PDFs\".

• plugin-factory

  Turn a plugin request into a tested plugin: scaffold, hooks, MCP server, tests, CI, validate.

• plugin-review

  Review someone else's plugin, skill or MCP server (folder, .zip, .skill, .plugin) and give one verdict.

• ponytail

  Lazy senior dev mode: simplest code that works, YAGNI, stdlib and native first; also reviews code for over-engineering and lists deferred shortcuts. Use when writing, fixing, refactoring or reviewing code, or for \"is this over-engineered\", \"what can we delete\", \"ponytail debt\".

• ratchet

  Fix and re-audit a target in rounds until two consecutive rounds find nothing new, each fix locked behind a test proven to fail without it. Use for \"audit this in rounds\", \"until there are no new findings\", or when work was declared done too early.

• repo-fix

  Review and fix a GitHub repository end to end: clone to a scratch folder, baseline the tests, fix on a branch with a failing test first; a draft PR only when asked. Use for \"fix this repo\", \"review this repository\", \"clean up this codebase\", \"fix the bugs in <repo>\".

• repo-review

  Review a repository, pull request or local folder you did not write, read-only and isolated: scratch folder, findings to local files. Use for \"review this PR\", \"look over this repo\", \"check this plugin folder\".

• research

  Tiered web research: low (1 searcher), medium (3 searchers, researcher writes), high (5, plus red-team and blue-verify). Use for low/medium/high research or \"research X\".

• setup

  Provision this machine for the workbench and keep it healthy: install, doctor, backup, profiles.

• skill-scan

  Security-scan installed Claude Code skills (or any folder) with NVIDIA SkillSpector and report per-skill risk. Use when a skill or plugin was just installed or written, or before trusting third-party skill code.

• spawn

  Action card for any Agent() call: pick the agent, then write a prompt the subagent can run without your context. Use before delegating anything, especially work that depends on project knowledge, hosts or money.

• ui-build

  Build, review or change a web UI the careful way: find the design system first, run the app locally, screenshot at 1440px and 390px in light and dark, run an accessibility check, fix. Use for \"build this page\", \"review this UI\", \"fix the layout\", \"make it responsive\", \"add dark mode\", \"po

• vdi-drive

  OPT-IN. Drive a remote desktop shown as an image in a browser (Citrix, Horizon, Azure Virtual Desktop, noVNC). Needs written authorisation.

• web-packager

  Check whether a skill or plugin can be uploaded to claude.ai and build the upload zip. Never uploads.

Subagents (11)

• blue-verify

  BLUE gate of a build/attack/verify run: verify a build, review a change before merge, break a red/blue/green tie, or judge a high-stakes one-shot where a wrong answer is expensive. Read-only verdict; if the answer is to BUILD something, hand it to flagship. Use as the gate, not the work. NOT for rou

• flagship

  Hard, ambiguous or long-horizon BUILD work that needs flagship-tier reasoning and effort: a design or migration with no obvious shape, a multi-hour build, an irreversible change, or a problem a worker already failed; also regulated work (legal, tax, medical, compliance) and final documents for a per

• green-team

  GREEN lane of a red/blue/green run: an independent alternative solution or clean-slate rebuild plan, on a different model from BLUE (Opus), so its mistakes don't correlate with BLUE's. Use alongside flagship/blue-verify (BLUE) and red-team (RED). NOT for attacking (red-team) or for final judging (bl

• implementer

  Build worker for one planned slice of a multi-slice run (2+ disjoint slices): implement exactly the owned files, run the slice's acceptance test, report the diff and the test output. NOT for diagnosis or planning (flagship), terminal or infrastructure slices (ops-runner), or review (blue-verify).

• ops-runner

  Server and application operations: health triage, Docker/Compose, host and OS admin, DNS, reverse proxies, TLS and firewalls, deploy pipelines and CI runners. Use for anything that runs on a box. Reads the infrastructure domain of the knowledge base for hosts before acting; this file names no host.

• red-team

  Adversarial pass: attack a plan, design, draft or finding and try to break it. Use for pre-mortems, refuting a blue-verify result, and the RED lane of any red/blue run. The FINAL verify of a run belongs to blue-verify, not this persona. NOT for building or fixing: this persona produces attacks, neve

• researcher

  Deep multi-source research: vendor and technology evaluation, competitive analysis, literature review, synthesis across many sources. Use proactively when a question needs more than two lookups. NOT for a single fact (inline), NOT for code search (use general-purpose with the code graph; the built-i

• scout

  Small-model reader for text or files HANDED to it: extract, classify, reformat, deduplicate, summarise or digest logs, transcripts, documents or search results into facts with locators, in a few tool calls. Makes no file edits. NOT for web research or tool loops (general-purpose), facts from memory,

• searcher

  Web searcher for ONE sub-question: searches the web, reads the pages and returns sourced facts. Use for the search lanes of the research skill (low, medium, high research), 'find every X' sweeps and lookups across many web sources. NOT for writing the report (researcher), tool loops that are not web

• security-auditor

  Security review: CVE and dependency risk, secret and credential exposure sweeps, infrastructure hardening, supply-chain review, OWASP issues, and the repo's own security scanners before a PR. Use proactively before shipping anything touching auth, secrets, or a public surface.

• sonnet-max

  Sonnet 5.5 at MAX effort. Use ONLY when the user explicitly asks for Sonnet at max effort; all other routing follows rules/model-policy.md.

Hooks (12)

• SessionStart

  ${user_config.python}

• SessionStart

  ${user_config.python}

• UserPromptSubmit

  ${user_config.python}

• UserPromptSubmit

  ${user_config.python}

• PreToolUse

  ${user_config.python}

• PostToolUse

  ${user_config.python}

• PostToolUse

  ${user_config.python}

• PostToolUseFailure

  ${user_config.python}

• PreCompact

  ${user_config.python}

• Stop

  ${user_config.python}

• ConfigChange

  ${user_config.python}

• InstructionsLoaded

  ${user_config.python}

Similar plugins