williamblair333
UserTurns AI sessions into compounding work — every session more correct, cheaper, and smarter than the last. Right > Cheap-in-total > Inventive > Local. Retrieval-first MCP stack for Claude Code, built on j.gravelle's tools.
Categories
Indexed Skills (40)
pre-mortem
Adversarial failure analysis across 12 dimensions before any consequential action — hooks, scripts, crons, config changes, deployments, gh pr create, git push. Blocks on HIGH/CATASTROPHIC until addressed.
prior-art-check
MemPalace search before any non-trivial task — "have we solved this before?" Fires before any coding, design, or architecture work; runs before the first substantive retrieval tool call.
session-end-checklist
Session wrap-up checklist — updates CHANGELOG, HANDOFF, considers ROADMAP/README, then commits. Invoke at session end or when the pre-commit hook blocks on missing docs.
smart-review
Auto-classifying code review router — rules floor + shadow classifier → dispatches code-review (low/medium/high) or adversarial-review (critical). Logs every classification to MemPalace for drift audit.
per-task-review-cycle
Implement → spec review → quality review → targeted fix → re-review cycle for multi-task subagent development. Use when executing an implementation plan where each task needs independent verification before moving to the next.
polling-bot-backlog-diagnosis
Diagnose and fix message backlog and notification-spam issues in cron-based polling bots — covers per-run deduplication flags and stale-message age filters
post-upgrade-mcp-integration
After any MCP stack upgrade, systematically discover new tools and integrate them into CLAUDE.md routing rules without being asked
session-status-briefing
Produce a comprehensive session-start status report combining git state, HANDOFF, code health metrics, risk surface, dead code candidates, and MemPalace HNSW health. Invoke at the start of any session to orient quickly without re-reading files.
stack-not-at-head-remediation
Remediate a HEALTHCHECK fail on stack-not-at-head — upgrade uv packages, re-index, pin embedding canary, run post-upgrade integration, update docs, commit, snapshot. Triggers whenever /health reports stack-not-at-head.
telegram-chat-history-persistence
Design and implement context continuity for a Telegram → Claude gateway. Use when the bot loses cross-session context or when choosing between transcript injection, SQLite buffering, RAG, and MemPalace for chat history.
auto-maintain-commit-and-deploy
Use when a cron-based auto-maintenance script commits new skills or plugins to git but the install script still has a hardcoded list that needs manual updating, or when new artifacts are committed but not immediately deployed (symlinked, copied, activated). Covers replacing hardcoded artifact lists with dynamic directory scanning and coupling git commits with immediate deployment steps.
install-script-cp-to-symlink
Use when a bash install script silently skips steps due to `cp` failing on pre-existing symlinks, or when converting copy-based installs to symlink-based for auto-propagation after `git pull`.
mempalace-fts5-malformed-index-repair
Repair a MemPalace SQLite FTS5 malformed inverted index reported by healthcheck as CRIT. Distinct from HNSW corruption (link_lists.bin) and HNSW 0-elements after reboot.
polling-bot-age-filter-fix
Apply a stale-message age gate to a Python Telegram polling bot to prevent a message backlog from exhausting the hourly rate-limit budget. Use after polling-bot-backlog-diagnosis confirms MAX_AGE_SECONDS is missing.
verify-pr-branch-before-resolve
Before resolving a PR's merge conflicts, verify which branch the PR actually targets — avoids wasted work on the wrong branch
agent-harness-backend-adoption-eval
Evaluate a proposed library or backend as a replacement for an existing one. Produces a structured assessment: architecture comparison, benchmark credibility, unproven risks, creative angles, and a staged recommendation. Use when a new library surfaces as a candidate to replace an embedded dependency (vector DB, cache, queue, etc.).
healthcheck-interactive-hints
Use when a shell healthcheck script uses hint() functions to point users at fix commands, and those commands silently fail or require a separate manual step, or when you want users to be prompted to run the fix immediately after a failure is reported.
judge
Before any Edit or Write lands, spawn an independent code-reviewer subagent that cross-checks the proposed change against structural evidence from jcodemunch (blast radius, changed symbols, untested symbols, PR risk) and serena (references). Catches hallucinated functions, wrong signatures, missed callers, and risky refactors before commit. Triggers when you're about to Edit or Write a non-trivial code change, or when the user asks you to verify/validate/double-check.
library-backend-adoption-eval
Structured evaluation of a new library as a potential backend replacement. Use when considering migrating from one backend (DB, index, vector store, cache) to another — covers architecture fit, performance claims, risks, creative experiments, and recommendation.
mcp-index-empty-diagnosis
Diagnose and fix an MCP retrieval tool (jdocmunch, jcodemunch, jdatamunch) that is registered and running but silently returns empty results due to an unpopulated index. Also wires in preventive measures so the gap can't recur silently.
mempalace-boot-repair-always-runs
Use when HNSW shows 0 elements at session start after a system reboot, a long-running repair process is already active, and the palace appeared healthy before shutdown. Distinct from link_lists.bin corruption — SQLite is intact, repair will succeed, but will repeat on every future reboot.
mempalace-dict-pickle-repair
Use when MemPalace healthcheck passes but `mempalace_search` fails with `'dict' object has no attribute 'dimensionality'`, or when dict-format pickles keep recurring after each repair run. Covers manual pickle migration, stdlib-only healthcheck probe design, and finding where the repair pipeline generates broken pickles.
mempalace-hnsw-corruption-fix
Diagnose and fix MemPalace HNSW index corruption where link_lists.bin grows to hundreds of GB due to chromadb-hnswlib 1.5.x type-confusion in Rust bindings. Use when link_lists.bin is abnormally large, mine crashes with OOM, or header.bin shows astronomical element counts.
mempalace-repair-mine-interference
Diagnose and resolve MemPalace repair failures caused by active mine jobs holding the SQLite write lock, an install script respawning miners, or FTS5 full-text index corruption discovered alongside HNSW corruption. Use when `mempalace repair` hangs, errors with "database is locked", or when FTS5 queries return nothing after HNSW is fixed.
orchestrator
Decompose a PRD or task into a structured JSON task manifest of parallelizable subtasks for multi-agent execution
platform-removal-cleanup
Use when dropping support for a platform (OS, runtime, environment) and need to scrub all artifacts: scripts, binaries, docs, template placeholders, config comments, and source code branches.
post-audit-mempalace-capture
After any adversarial/red-team analysis of a component, skill, or control mechanism, write two durable MemPalace entries — design invariants + audit baseline, and closed attack vectors — so future sessions can answer "is this change a regression?" without relying on git history alone.
readme-sync
Audit a README against actual repo contents, identify undocumented features, and make targeted edits to bring docs in sync with what's actually installed/shipped.
review-queue-triage
Audit _review/ items, classify by completion signals (status: resolved, REPO-ASSESSMENT + landing commit), auto-move resolved items to _reviewed/, leave uncertain items for user decision.
stop-hook-dedup-guard
Use when a Claude Code Stop hook sends a Telegram (or other) notification and the same message arrives duplicated at nearly the same timestamp. Symptom is two identical "⏳ Running…" + message pairs appearing within seconds of each other after closing Claude Code.
telegram-gateway-security-audit
Diagnose and harden a Telegram → Claude CLI gateway: fix update_id deduplication failures that spawn duplicate sessions, and inject a disclosure-restriction system prompt so the bot never leaks OS/infra details over the channel.
telegram-inline-button-promote
Wire up a Telegram inline keyboard button in a polling bot — add button to outbound notification, subscribe to callback_query updates, answer the callback, and collapse a multi-step confirm flow into a single button press.
token-economy-prompt-authoring
Craft a minimal, effective system prompt to constrain Claude verbosity in plain web sessions (no harness tooling). Use when you need to paste a behavioral contract into a new Claude.ai conversation or share one for a context where superpowers/skills aren't available.
validate-external-audit
Use when responding to an external critique, audit, or code review of a codebase before accepting or rejecting any finding.
dream-synthesizer
Analyze Claude Code session traces and extract recurring mistakes and proven playbooks for future sessions
gemini-auto-skill
Monitors Gemini CLI sessions for reusable workflows and drafts native Gemini skills.
community-pr-stakeholder-response
Use when tagged as a stakeholder on an open-source PR or issue comment, especially when the PR has evolved since your last review, involves architectural changes, or requests sign-off on a change touching shared dependencies or parallel work.
outcomes
Rubric-aware grader agent — evaluates working agent output against a rubric in a fresh context and produces a structured gap report
stale-pending-memory-guard
Before reporting any memory hit containing "pending/awaiting/needs/consider" as current fact, verify it against the actual source. Prevents propagating resolved items that memory hasn't caught up to.
terse-reply
Edit the most recent draft reply to strip verbosity. Use when the user invokes this skill or says "cut the bullshit", "tighten this", "too verbose", or similar.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.