Menu
How to clean up duplicate/untracked skills in the skills repo
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Verify a code change actually works by building/running the app and observing it at its real surface (CLI, API, UI, library, agent), capturing runtime evidence rather than trusting tests. Make sure to use this skill whenever the user has changed code and wants to know it works, is about to merge/push and wants confidence, says "did this actually work", "verify this works", "prove it works", "confirm the change", "make sure it works", wants runtime evidence, or is re-running tests / importing-and-calling just to check behavior, even if they never say the word "verify". When in doubt after any code change, reach for this. For post-deploy production health checks use verify-deploy; for static correctness/quality review use simplify or code-review.
Test an interactive lesson/course (or any "instructions to an AI" skill) by self-play. An agent plays BOTH the instructor following the lesson script AND a calibrated student persona, producing full turn-by-turn transcripts of every lesson, then publishes the raw transcripts to a single static page. Use when asked to "run lesson transcripts", "test the course end to end", "self-play the lessons", "publish raw test transcripts", "walk a synthetic student through every lesson", or to QA an interactive-instruction skill by actually running it rather than just reviewing findings. Distinct from dogfood and adversarial-ux-test (web-app browser QA) and synthetic-userstudies (findings plus a few cherry-picked transcripts). This one captures the COMPLETE run of every lesson and ships them all raw.
Daily memory garbage collection for MEMORY.md / USER.md. Apply decay rules, drain .pending.md, consolidate near-duplicates, maintain canonical theme tags, prune old episode and session files. Invoke when asked to "run memory GC", "clean up memory", "apply memory decay", or from the scheduled cron job.
Monitor a PR until it's ready to merge. Watches CI, reads reviews, checks scope, fixes blocking issues, opens follow-up PRs for low-priority comments, and repeats. Use when: babysit this PR, watch this PR, monitor PR, fix and watch PR, keep this PR green.
Review code changes by analyzing git diffs, leaving inline comments on PRs, and performing thorough pre-push review. Works with the gh CLI or falls back to git + GitHub REST API via curl. Use when: review this PR, review a pull request, look at this PR, code review, review my changes before I push, pre-push review, leave inline comments on a pull request.
Run iterative self-referential development loops using the Ralph Wiggum technique. Use when tasks need repeated iteration, TDD cycles, greenfield builds, or autonomous refinement until tests pass or completion criteria are met. Triggers on ralph loop, ralph mode, iterative loop, autonomous loop, /goal.
| name | skills-cleanup |
| description | How to clean up duplicate/untracked skills in the skills repo |
| tags | ["skills","maintenance","cleanup"] |
~/.hermes/skills/ (git: $SKILLS_REPO)creation_nudge_interval: 0 in config.yaml)~/.hermes/skills/ and skills.external_dirs (currently empty) — nothing else~/.hermes/skills-archive/~/.hermes/skills/<category>/<name>/git status) against remote (git ls-tree -r --name-only origin/main)~/.hermes/skills-archive/ (recoverable)Do not do cleanup or PR work directly in the runtime checkout (~/.hermes/skills). Preserve first, then move the review work into a project worktree:
cd ~/projects/skills
git fetch origin main
BRANCH="wip/preserve-skills-$(date +%Y%m%d)"
git worktree add ~/projects/_worktrees/$BRANCH -b "$BRANCH" origin/main
Copy or patch only the intentional skill changes into the worktree. Exclude generated/runtime state (.usage.json, .usage.json.lock, .curator_state, .curator_backups/, temporary reports, moved-file markers). Before committing, inspect git diff --name-status and stage explicit files rather than blanket-adding the runtime checkout.
For longer cleanup/review checklists, see references/runtime-skills-preservation.md and references/runtime-snapshot-pr-review.md.
Before pushing, scan for violations (this repo is PUBLIC):
grep -rnE '(personal-email@example\.com|PRIVATE_DOMAIN|ACCOUNT_ID|srv-[a-z0-9]{10,}|evg-[a-z0-9]{10,}|AuthKey_[A-Z0-9]+|\+1[0-9]{10}|password in)' --include="*.md" --include="*.py" --include="*.sh" .
Critical rules:
references/ file and inline into SKILL.md if result exceeds 500 lines. Restore the reference file.$VAR_NAME for all account-specific values. See ../../coding/babysit-pr/references/public-repo-redaction.md for the full list.github/, autonomous-ai-agents/, software-development/, mlops/, media/, mcp/, leisure/, red-teaming/, note-taking/, email/, smart-home/, gaming/, analytics/, diagramming/, domain/, feeds/, gifs/, inference-sh/, dogfood/
Skills templated from a client project often have dead examples baked in (wrong audience, irrelevant copy, client-specific search queries). These mislead creative generation. Load references/inherited-client-examples.md for detection, cleanup rules, and a worked example.
When the user says the repo has "too many top-level folders" or categories feel off, load references/category-reorganization.md for the full audit and move playbook. Covers bloat detection, merging overlapping categories, moving project-specific skills to internal/, and keeping README.md and CLAUDE.md in sync.
When adding, removing, or renaming skills, update the README.md category table with accurate skill counts.