Hefrock

agent-eval

Designs and runs evaluations for LLM or agent outputs — builds rubrics, sets up LLM-as-judge scoring, creates regression test sets, and reports pass rates with concrete failure examples. Use this skill whenever the user wants to evaluate, test, grade, or score an agent's or LLM's outputs; asks "how do I know if this is working," "is this any good," "set up an eval," or "did this prompt change make things worse"; needs a rubric for judging quality; wants to compare two prompts, models, or outputs; or wants to catch regressions before shipping a change. Also trigger when reviewing agent trajectories specifically — did the agent pick the right tool, the right arguments, the right sequence — not just the final output.

broadcast

Produces a daily healthcare AI news audio briefing — ingests from eleven registered sources (PubMed, arXiv, medRxiv, FDA guidance, regulations.gov, FDA MAUDE device adverse events, ONC/ASTP, CMS, and industry RSS feeds), dedupes and ranks stories, pins every claim to its source via a real evidence-pinning MCP server, generates a spoken script, rewrites each story with AI narration (grounded, with an automatic fallback to plain text if the narration can't be verified against its source), runs a QA gate, synthesizes audio via Gemini TTS, and produces GitHub-Pages-ready distribution artifacts (podcast RSS feed + Obsidian vault note). Use this skill when the user wants to run, generate, produce, or publish a healthcare AI briefing episode or podcast; asks about the status of a broadcast episode or run; wants to debug a failed or partial episode (QA failures, narration failures, TTS rate limits); wants to add/remove/tune an ingest source; or asks how the broadcast pipeline works. Requires GEMINI_API_KEY and a loca

deid-reid-harness

Build and run an adversarial de-identification ⟷ re-identification evaluation harness for clinical text: generate synthetic clinical notes with ground-truth identifier spans, run a de-id pipeline (defender) over them, and score leakage against an attacker. Use this whenever the task involves evaluating a PHI scrubber or de-id pipeline, measuring re-identification risk, testing HIPAA Safe Harbor coverage or Expert Determination risk, building a privacy-utility frontier for clinical NLP, or comparing de-identification approaches — even if the user only says "test my de-identifier," "how leaky is this scrubber," or "de-id eval." Prefer this over an ad-hoc script whenever synthetic clinical data, PHI, or de-identification evaluation is mentioned.

issue-reconciler

Finds GitHub issues that are actually already resolved but were never closed — either because a merged PR said "Closes

privacy-linter

Deterministic pre-disclosure privacy scanner for git-staged changes — flags Direct PII (email, phone, SSN, Luhn-valid credit card, IP address), Secrets (AWS/GitHub/Slack/Stripe/Google/Anthropic tokens, private key blocks, hardcoded-credential assignments), and Metadata risk (image/document file types that commonly carry EXIF/embedded properties, with a confirmed EXIF GPS check when Pillow is available) before a commit lands. `--scan-history` walks the full commit history for PII/secrets ever introduced, even if later removed from HEAD. Can also strip EXIF metadata from a JPEG/TIFF outright with `--strip-metadata`, and track findings over time with `--log-dir` + `scan_log_history.py` (bridges into agent-eval's score_eval.py for trend/regression reporting, same pattern broadcast's qa_gate_history.py uses). Runs entirely locally with no model or network call — pattern-matching only. Use when the user wants to check a commit, diff, or file for leaked PII or secrets before it's committed or shared, asks "will this

privacy-threat-oracle

Deterministic, rule-based privacy decision engine — evaluates a proposed action (which identity/compartment it's taken from, who can see it, what sensitive content it contains) against a threat model of adversary classes and identity compartments, and reports a recommendation (proceed / proceed_with_modification / decline) with a stated residual risk and reason. Its five-class content-sensitivity vocabulary (direct_pii, secret, metadata, inference_cue, stylometric) matches privacy-linter's own leak_class values for the first three, so privacy-linter's --json output plugs in directly as the content-sensitivity input via --from-linter-json; stylometric matches style-obfuscator's --emit-findings bridge instead, and inference_cue is a reserved class neither tool currently produces (see "How this works" for the full breakdown). Runs entirely locally, no model or network call — a rule table over a fixed schema, not an LLM judgment call. Use when the user wants to check whether sharing/posting something crosses an i

repo-pincer

Reverse-engineers a codebase by reconciling what it claims to do against what it actually does. Runs two independent passes — top-down from docs/README/manifests/API surface, bottom-up from entry points through the actual implementation — then classifies every claim as Confirmed, Drift (was true, isn't anymore), Aspirational (documented but never built), or Silent (implemented but undocumented). The discrepancy list is the deliverable, not either summary alone. Use this when onboarding onto an unfamiliar codebase, auditing whether documentation matches implementation, evaluating a dependency or vendor repo before committing to it, or when a long-lived project's docs and code have drifted apart. Triggers on "reverse engineer this repo," "does the documentation match the code," "audit this codebase," "onboard me onto this project," or "find where the docs and code disagree." Runs standalone — no MCP or wiki setup required. If an obsidian-vault MCP server is connected, offers to compile its output into Sources/r

research-ledger

Captures a deep-research run's full claim ledger — every claim extracted, its verdict, and whatever vote/reasoning detail the run's own output exposes — as a structured JSON artifact, before that detail is lost to synthesis or context compaction. Hands the ledger to wiki-warehouse for durable, content-hash-addressed storage, so a research run's rejected claims and per-source detail survive even though only the confirmed subset gets promoted into Knowledge/ pages. Use this immediately after any deep-research skill run completes, before its findings are synthesized into the vault. Triggers on "save the claims ledger," "archive this research run," "warehouse the research," and the command /research-ledger. Requires the obsidian-vault MCP server and wiki-warehouse's prerequisites (the knowledge-warehouse repo cloned locally).

style-obfuscator

Computes a deterministic stylometric fingerprint of a piece of writing — function-word frequency, sentence-length distribution, punctuation habits (em dash rate, contraction rate), approximate passive-voice rate, and repeated-phrase detection — and optionally scores its similarity against a reference corpus of your own known writing (your vault, past posts). --corpus pools many of your own pseudonymous drafts into one aggregate fingerprint and reports the "aggregation gap" against a single-document check — catching the case where individually-clean posts are collectively identifying, the way real cluster-based stylometric matching actually works. No model or network call; classic authorship-attribution stylometry (Mosteller & Wallace, Burrows' Delta) has always been word-frequency statistics, not semantic understanding. Use when the user wants to check whether a draft (a GitHub PR/README/commit message, a LinkedIn post, any public professional writing) carries a distinctive fingerprint that could later be use

wiki-governor

Runs the wiki's self-governing maintenance loop and holds it accountable to its own constitution. Orchestrates wiki-librarian (structure), wiki-synthesizer (compilation), and wiki-warehouse (cold-storage integrity, if in use), then adds the things none of them do — a constitution-compliance audit, a tracked wiki health score, and a knowledge-gap queue. Use on a weekly cadence or when the vault hasn't been maintained in a while. Triggers on "govern my wiki," "maintain my wiki," "is my wiki healthy," "wiki health score," "check constitution compliance," "what am I missing in my vault," and the command /govern. Requires the obsidian-vault MCP server connected. Pairs with wiki-librarian, wiki-synthesizer, and wiki-warehouse, which it invokes rather than reimplements.

wiki-librarian

Maintains the structural health of the wiki. Audits for broken links, orphan pages, stale notes, near-duplicates, and contradictions. Proposes and applies fixes with user confirmation. Run weekly or before a major synthesis session. Triggers on "audit my wiki," "clean up my vault," "check wiki health," "find duplicates," "fix broken links," and the command /audit. Requires the obsidian-vault MCP server connected. Pairs with wiki-operator for on-demand edits and wiki-synthesizer for content compilation.

wiki-privacy-audit

Audits an Obsidian vault's notes for Direct PII and Secrets accidentally captured while journaling or researching — reuses privacy-linter's deterministic scan_diff.py rather than reimplementing detection. Reports findings and proposes redaction/removal with confirmation; never edits vault content automatically. Use when the user wants to check their notes/journal for leaked personal information or credentials, asks "did I ever paste something sensitive into my vault," or wants a periodic privacy health check on their knowledge base. Triggers on "audit my vault for PII," "check my notes for secrets," "did I leak anything into my vault," "privacy audit my wiki," and the command /privacy-audit. Requires the obsidian-vault MCP server connected (to resolve the vault's local filesystem path) and privacy-linter's scan_diff.py (a sibling skill, invoked directly as a subprocess, not reimplemented). Does not cover attached-image metadata or vault git history — see "What's NOT covered here."

wiki-synthesizer

Compiles raw material into the wiki. Automatically preprocesses unstructured journal entries, then promotes ideas into Knowledge/ concept pages and compiles Sources/raw/ into source pages. Updates Maps/_context.md after each run. Use this after learning sessions. Triggers on "synthesize my notes," "compile my wiki," "promote from journal," "process raw sources," and the command /synthesize. Requires the obsidian-vault MCP server connected. Pairs with wiki-operator for on-demand edits and wiki-librarian for maintenance.

wiki-teacher

Portfolio-aware project accountability for several concurrent projects — /checkin, a forcing function auto-suggested at session start. Surfaces 1–2 projects that genuinely need attention (never a dump of everything overdue), batch-elicits priority when it's needed instead of asking one per day, and answers "what should I work on" even when nothing's overdue. State lives declaratively in each project note's own frontmatter (priority, checkin_interval, status) — no separate tracking mechanism. Triggers on "check in on my projects," "what should I be working on," and the command /checkin. Requires the obsidian-vault MCP server connected. Deliberately narrow for now — see the note below.

wiki-warehouse

Ingests raw research documents (PDFs, ebooks, scans, articles) into a separate private "cold storage" GitHub repo, then writes a lean pointer note into the Obsidian vault — keeping originals and full extracted text out of the vault so it stays fast and git-diffable. Use when the user wants to file, archive, warehouse, or "add a document/paper/PDF to my knowledge base," or to keep large source files outside the vault. Triggers on "warehouse this," "ingest this document," "add this PDF to cold storage," "file this paper," and the commands /ingest and /warehouse-audit. Requires the obsidian-vault MCP server connected and the knowledge-warehouse repo cloned locally. Pairs with wiki-synthesizer (which compiles the resulting Source note into concepts) and wiki-librarian (structural audits).

my-skill-name

A clear, specific description of what this skill does and exactly when Claude should use it. Be concrete about trigger conditions — what the user says or what task context should cause this skill to load.