Agent Skills

A personal collection of Agent Skills — self-contained capabilities that any compatible AI agent can load on demand. Skills are built on the open standard Anthropic originally published, now supported by Claude, Codex CLI, Gemini CLI, GitHub Copilot, Cursor, and 25+ other platforms, so nothing here is locked to one tool.

Each skill is just a folder: a SKILL.md file with instructions, plus scripts/references/assets where needed. See CONTRIBUTING.md if you want to add one or keep it portable across platforms.

Important

Stalled-work tracking. Tasks blocked on someone else (a signature, a sign-up, a records request) are tracked as GitHub issues and surfaced weekly — see docs/stalled-work-tracking.md.

Jump to: Skills · Installing · Repo structure · Wiki system

Skills

Agent design — build and evaluate other agents

Skill	What it does
agent-eval	Turns "does this agent actually work?" into a repeatable score: rubrics, LLM-as-judge grading, and regression sets that catch when a prompt change made things worse.
agent-redteam	Generates adversarial test cases to check that an agent fails safely — refuses, hedges, or degrades gracefully — instead of confidently getting it wrong. Pairs with agent-eval for scoring.
repo-pincer	Reverse-engineers a codebase: reads what the docs claim, reads what the code actually does, and reports exactly where they disagree.
issue-reconciler	Finds GitHub issues that are actually already resolved but never got closed — a merged PR that said "Closes #N" but didn't take effect, or a skill that quietly got built with nobody linking the PR. Proposes closures with evidence; never closes anything on its own.

Privacy — catch leaks before they happen

Skill	What it does
privacy-linter	Scans a git diff before you commit for leaked emails, SSNs, credit cards, API keys, and photo GPS metadata — entirely offline, no model call. Can strip EXIF metadata outright, dig through commit history for old leaks, track your leak rate over time, and block a commit on high-severity findings (the installed pre-commit hook does this by default).
deid-reid-harness	Stress-tests a clinical de-identification pipeline by trying to re-identify the patients afterward, across three attack types plus a privacy/utility tradeoff score. Runs offline with real statistical confidence intervals. Moved here from Agent design — once real DUA-governed patient data is in scope (see the open safety-hardening issue), this is part of the privacy-sensitive data chain, not just an eval harness that happens to score PHI. (sample results)
privacy-threat-oracle	A rule-based second opinion on "is it safe to share this?" — weighs which identity of yours it's coming from, who could actually see it, and what's in it against a threat model of real adversaries, and gives a clear proceed / modify / decline. Plugs directly into privacy-linter's output.
style-obfuscator	Computes a stylometric fingerprint of a draft (function-word frequency, punctuation habits, sentence stats) and scores its similarity against your own known writing — so you know what a matcher would key on before posting something you don't want traced back to you. Flags only; never rewrites.

Knowledge management — a personal wiki that maintains itself

Skill	What it does
wiki-operator	Your everyday interface to the vault — /learn, /update, /connect, /ask, /review, /quiz, /map, /source, /clean, /health.
wiki-synthesizer	Turns raw journal entries and saved sources into proper, linked knowledge pages. Run after a learning session.
wiki-privacy-audit	Runs privacy-linter across your entire Obsidian vault instead of one diff, so a stray SSN or API key pasted into a journal entry doesn't sit there forever. Has a wrapper for unattended/scheduled runs.
wiki-librarian	Weekly housekeeping — finds broken links, orphaned notes, duplicates, and contradictions, and proposes fixes for you to confirm.
wiki-governor	Runs the librarian, synthesizer, and warehouse on a schedule, then grades the vault's own health against its own rules and tracks a knowledge-gap queue.
wiki-teacher	A weekly /checkin that looks across several ongoing projects and surfaces the 1-2 that genuinely need your attention — never a dump of everything at once.
wiki-warehouse	Archives PDFs, ebooks, and scans in a separate private repo, leaving only a lightweight pointer note in the vault so it stays fast.
research-ledger	Saves a deep-research run's full receipts — every claim and how it was verified — before that detail gets compressed away during synthesis.

Content automation

Skill	What it does
broadcast	Produces a daily healthcare-AI news audio briefing: pulls from eleven registered sources, dedupes and ranks stories, pins every claim to its source through a real evidence-pinning MCP server, writes a script, synthesizes audio, and publishes a podcast feed + vault note. Requires GEMINI_API_KEY and a locally-built evidence-pinning-mcp server.

Most knowledge-management and privacy-vault skills require the obsidian-vault MCP server — see Wiki system below for setup. Full flag-by-flag documentation for any skill lives in its own SKILL.md, not here.

Installing a skill

Claude Code (plugin marketplace):

/plugin marketplace add Hefrock/agent-skills
/plugin install agent-eval@hefrock-agent-skills

Install any other skill the same way — swap agent-eval for the plugin name from the table above.

Getting updates after this repo changes:

/plugin marketplace update hefrock-agent-skills

This pulls the latest manifest, but an already-installed skill only picks up a content change (an edited SKILL.md, a bug fix) once .claude-plugin/marketplace.json's shared metadata.version has been bumped past what you have. It is not automatic just because the file changed on GitHub — if a skill's behavior doesn't match what its SKILL.md currently says, check whether the version was actually bumped for that change before assuming the update didn't take.

A plugin new to the marketplace still needs its own /plugin install <name>@hefrock-agent-skills afterward. If it comes back "not found" right after updating, restart the CLI session and retry.

Claude Code (manual, no plugin system):

git pull origin main   # if you already have the repo cloned
cp -r skills/agent-eval ~/.claude/skills/

claude.ai: Zip the individual skill folder (e.g. skills/agent-eval/) and upload via Settings → Features → Custom Skills (requires a paid plan with code execution enabled).

Other platforms (Codex, Gemini CLI, Cursor, etc.): Copy the skill folder into whatever directory that platform scans for skills — the SKILL.md format works unmodified.

Repo structure

Expand directory tree

agent-skills/
├── .github/
│   ├── workflows/ci.yml        # builds + tests the MCP server, runs the Python skill suites
│   └── ISSUE_TEMPLATE/         # blocked-human / dated-followup templates for stalled-work tracking
├── .claude-plugin/
│   └── marketplace.json        # Claude Code-only install metadata — optional, additive
├── docs/
│   └── stalled-work-tracking.md # how the blocked-human / dated-followup convention works
├── skills/                     # flat — one folder per skill, no category nesting
│   ├── broadcast/              # daily healthcare AI audio briefing pipeline — scripts/orchestrate.py, 563-test suite
│   ├── agent-eval/              # rubric-based evals, LLM-as-judge, regression test sets
│   ├── agent-redteam/           # adversarial case generation, pairs with agent-eval
│   ├── deid-reid-harness/       # clinical de-id/re-id eval — scripts, refs, 72-test suite
│   ├── repo-pincer/             # codebase reverse-engineering — claims vs. reality reconciliation, scripts/check_structural_claims.py + track_findings.py, test suite
│   ├── issue-reconciler/        # GitHub issues vs. reality reconciliation — scripts/find_closing_references.py, test suite
│   ├── privacy-linter/          # pre-disclosure PII/secrets/metadata scanner — scripts/scan_diff.py, test suite
│   ├── privacy-threat-oracle/   # rule-based compartment/adversary decision engine — scripts/oracle.py, test suite
│   ├── style-obfuscator/        # stylometric fingerprint + reference-corpus similarity — scripts/fingerprint.py, test suite
│   ├── wiki-operator/           # on-demand vault operations
│   ├── wiki-synthesizer/        # journal preprocessing + concept page compilation
│   ├── wiki-librarian/          # structural health audits — scripts/check_vault.py, 28-test regression suite
│   ├── wiki-governor/           # maintenance loop + compliance + health score — scripts/health_score.py, 18-test regression suite
│   ├── wiki-teacher/            # /checkin (project accountability) — scripts/wiki_teacher.py, 37-test regression suite
│   ├── wiki-warehouse/          # raw-document cold storage (external repo) + vault pointers
│   ├── wiki-privacy-audit/      # vault-wide PII/secret audit — reuses privacy-linter's scanner
│   └── research-ledger/         # deep-research claim ledger, warehoused ahead of synthesis
├── mcp/
│   ├── evidence-pinning/       # MCP server required by broadcast — durable claim/source provenance log
│   └── obsidian-vault/         # MCP server required by wiki-operator
│       ├── src/index.ts        # 10 tools: search, read, write, append, patch, query, links, delete
│       ├── test/               # end-to-end STDIO tests — `npm test`
│       └── README.md           # setup and configuration guide
├── knowledge-os/
│   ├── constitution.md         # 10 laws Claude follows when operating the wiki
│   ├── architecture.md         # component map, data flow, note lifecycle
│   └── sitrep.md               # living status + gap analysis for the wiki system
├── templates/                  # note templates copied into vault by setup-vault.sh
│   ├── concept.md
│   ├── journal.md
│   ├── source.md
│   └── map.md
├── bin/
│   └── setup-vault.sh          # one-command vault bootstrap (creates folders, copies templates + constitution)
├── skill-template/              # starting point for a new skill
└── CONTRIBUTING.md             # how to add a skill, including portability rules

Wiki system

The wiki skills (wiki-operator, wiki-synthesizer, wiki-librarian, wiki-governor, wiki-teacher, wiki-warehouse, wiki-privacy-audit) form a complete personal knowledge system built around an Obsidian vault:

• wiki-warehouse adds a separate private repo for raw documents, so originals stay out of the vault while still being indexed by a content-hash pointer.
• wiki-teacher adds project accountability on top of your project portfolio — orthogonal to the rest, which are about the knowledge graph itself.
• wiki-privacy-audit pairs with privacy-linter instead of reinventing detection — it's the one wiki skill whose "audit" is fully mechanical, so it runs the linter's scanner directly against vault notes rather than reasoning over each one.
• research-ledger slots in just ahead of wiki-synthesizer — it warehouses a deep-research run's full claim ledger (including what got rejected) before synthesis only promotes the confirmed subset into Knowledge/.

Every wiki skill requires the obsidian-vault MCP server connected — nothing works without it. To enable it:

1. Build it once:

   cd mcp/obsidian-vault && npm install && npm run build

2. Point it at your vault. Three ways, easiest first:
   • Claude Code CLI (recommended):

     claude mcp add obsidian-vault \
       -s user \
       -e OBSIDIAN_VAULT_PATH=/absolute/path/to/your/vault \
       -- node /absolute/path/to/agent-skills/mcp/obsidian-vault/dist/index.js

     -s user registers it at the user level — available in every project, which is how this server is meant to run — and edits ~/.claude.json for you, avoiding the occasional weirdness of hand-editing that file while a running Claude Code process has it open.
   • setup-vault.sh — run ./bin/setup-vault.sh ~/path/to/vault; it bootstraps the folder structure, copies templates and the constitution into System/, and prints a config snippet with your paths filled in, for the manual route below.
   • Manual — add this to ~/.claude.json yourself:

     {
       "mcpServers": {
         "obsidian-vault": {
           "command": "node",
           "args": ["/absolute/path/to/agent-skills/mcp/obsidian-vault/dist/index.js"],
           "env": { "OBSIDIAN_VAULT_PATH": "/absolute/path/to/your/vault" }
         }
       }
     }

3. Restart Claude Code and verify: run /mcp — expect obsidian-vault connected with 10 tools. Full tool reference: mcp/obsidian-vault/README.md.

This is a per-device setup step — the server is a local process, so a new machine needs its own build and its own ~/.claude.json entry (with paths for that machine), even if it's pointed at the same synced vault.

The operating layer lives in this repo:

Path	Purpose
knowledge-os/constitution.md	10 non-negotiable rules Claude follows when operating the wiki
knowledge-os/architecture.md	Component map, data flow, and note lifecycle reference
knowledge-os/sitrep.md	Living status + gap analysis for the wiki system — what's shipped, what's untested, what's next
templates/	Note templates (concept, journal, source, map) — copied into System/templates/ in your vault by setup-vault.sh

License

See LICENSE.