Skills repo

Dan's personal Claude Code skills, public on GitHub as chendaniely/skills. Each top-level folder is one skill. Open work lives in TODO.md and is mirrored as GitHub issues.

Install and layout

• Install: ./install.sh symlinks each <skill>/ into ~/.claude/skills/<skill>. Edits here apply live, even in running sessions. Re-run it after adding or renaming a skill.
  • It never touches entries in ~/.claude/skills that aren't its own links, such as skill folders other tools installed there.
  • If a skill's name is already taken there by a folder or a different link, it prints skipped and leaves it.
  • It removes its own links whose skill folder no longer exists.
• Keep skills flat at the top level (decided 2026-09-16). Claude Code only discovers personal skills one level deep (~/.claude/skills/<name>/SKILL.md), so group folders wouldn't group anything, and a move breaks the paths listed below.
• Marketplace (added 2026-09-23): .claude-plugin/marketplace.json also publishes the skills as five plugins (vault, wardrobe, teaching, running, coding) for /plugin marketplace add chendaniely/skills. A new or renamed skill must be listed there too.
  • Plugin installs prefix skill names (/wardrobe:ootd) and are cached copies, so edits apply only after a push and plugin update. Anything that loads a skill by bare name (like the daily-email-digest task) needs the symlink install.
  • Don't use both installs on one machine, or every skill loads twice.

Paths other repos depend on

When a skill folder is renamed or moved, update all of these in the same sitting:

• The Obsidian vault's CLAUDE.md: it points, by this checkout's absolute path (print it with git rev-parse --show-toplevel, then search the vault's CLAUDE.md for it), to product-search, clothing-inventory, ootd and capture-conversation, and to the repo root.
  • Find the vault through Obsidian's vault list (~/Library/Application Support/obsidian/obsidian.json); it is the vault named vault.
  • A safety hook may block broad commands over the whole vault. Use narrow paths or the vault's own scripts (scripts/lint.py).
• ~/git/private/claude-cowork: its CLAUDE.md and README.md point to email-digest/SKILL.md and say to commit skill changes here.
• The daily-email-digest scheduled task (~/.claude/scheduled-tasks/daily-email-digest/SKILL.md): it loads email-digest by name and cites its path here. It stops if the skill can't be loaded, so a rename ends the daily digest until the task is updated.
• Course repositories using class-notes: each one's Makefile runs ~/.claude/skills/class-notes/scripts/class_notes.py (the symlink install; with the plugin install, pass CLASS_NOTES=<path>), and its README-prompt.md tells students to clone this repo and run class-notes/scripts/class_notes.py. Renaming the skill, its scripts/ folder or the script breaks both. Find them with grep -l class_notes.py over the course repos' Makefiles.
• chendaniely/genai-learning-moments: its README.md, CLAUDE.md and about.qmd link to capture-learning-moment by its GitHub path, and its CLAUDE.md assumes the skill defers to it.
• The symlinks: re-run ./install.sh.

How the skills fit together

• email-digest is driven by the private claude-cowork repo. Its CLI subcommands, JSON shapes and checklist markers must match that repo, and claude-cowork's CLAUDE.md says the same from its side.
  • Changes usually come from a claude-cowork session, so check git status before touching this skill.
  • Unattended runs must start in claude-cowork, because only its machine-local settings pre-approve the digest commands.
• catch-up and email-digest write the same daily note. digest.py render puts its ## Email checklist between <!-- ea-email:begin todos … --> and <!-- ea-email:end todos -->, right after ## Today, and the begin marker comes before the ## Email heading. catch-up ends its ## Today rewrite at that marker and never touches lines inside the block. The vault's CLAUDE.md (## Conventions) requires the same.
• clothing-inventory, ootd and product-search refer to each other by name. ootd runs clothing-inventory as a required sub-skill, so rules that touch both (size, aliases, listing line, photos) must agree.
• Vault skills follow the vault's conventions, and the vault's CLAUDE.md wins.
  • catch-up, capture-conversation and ootd look up folder names in the vault's CLAUDE.md.
  • clothing-inventory hard-codes this vault's inventory/ paths and defers to inventory/Inventory.md.
  • email-digest gets its vault settings from claude-cowork's config.
  • Frontmatter changes wait for the vault's Phase 10 migration (see TODO.md).
• Inbox and transcripts:
  • catch-up files inbox/ notes, and inbox/ is also where ootd looks for unlogged outfit photos.
  • Plaud transcripts reach the vault's sources/transcripts/ through the vault's audio pipeline. catch-up never files them, and capture-conversation should link them instead of transcribing again (see TODO.md).
• Standalone: capture-learning-moment (writes to $LEARNING_MOMENTS_DIR, else ~/learning-moments/), creating-ultra-crew-guides and external-app-review don't use the vault.
  • external-app-review writes into the reviewed app's own repo, and its two scripts need Playwright resolvable from that repo (or NODE_PATH) plus Google Chrome.
  • When that folder is a Quarto website (it has _quarto.yml), capture-learning-moment follows the site's own CLAUDE.md for the post folder and conventions. Dan's is the public chendaniely/genai-learning-moments, whose posts live in posts/.
• class-notes doesn't use the vault either: it works inside a course repository, which holds prompt-notes.md (the prompt, sent whole), README-prompt.md (student docs) and one folder per class session.
  • Class recordings reach course repos through class-notes fetch, separately from the vault's audio pipeline.
  • It uses the Plaud CLI (@plaud-ai/cli), which has its own login, not the Plaud MCP. After a CLI upgrade, run fetch --verify on an existing session (see references/plaud.md), since the CLI's text output is parsed with regexes.
  • templates/ holds the boilerplate for a new course. Keep it free of course-specific details, and of real hostnames and paths.

Editing a skill

• Frontmatter:
  • name equals the folder name.
  • description is at most 1024 characters and says what the skill does and when to use it.
  • Single-quote the description, doubling any internal apostrophes, if it contains # or : . An unquoted #todo once silently cut catch-up's description short.
  • Check every skill at once. This fails loudly on a name/folder mismatch or an over-long description, and prints each file's line count for the 500-line limit:

    for f in */SKILL.md; do uv run -q --with pyyaml python -c "import yaml,re,sys; p=sys.argv[1]; t=open(p).read(); fm=yaml.safe_load(re.match(r'^---\n(.*?)\n---\n',t,re.S).group(1)); assert fm['name']==p.split('/')[0], p+': name != folder'; assert len(fm['description'])<=1024, p+': description too long'; print(fm['name'], len(fm['description']), t.count(chr(10)), 'lines')" "$f"; done

• Length: keep SKILL.md under 500 lines. Move detail that matters only sometimes into references/, one level deep and linked from SKILL.md. Record new lessons as short rules there, not as dated stories in the steps.
• Side effects: a skill that moves or rewrites files triggers only on an explicit request. catch-up runs only when asked. ootd offers first when it finds a photo mid-task.
• Self-edits: no skill edits the vault's CLAUDE.md or its own SKILL.md on its own initiative. It proposes the change and applies it after Dan says yes.
• grep in Claude Code sessions prints recursive paths without a leading ./. Write path filters that accept both forms, for example grep -vE '^(\./)?TODO\.md:'.
• Tests:
  • ootd has the only regression test: ootd/tests/prompt.md (the prompt) and ootd/tests/check_green.py (the grader, with the fixture recipe in its header).
    • Re-run GREEN after any change to ootd or to clothing-inventory, which ootd runs as a sub-skill.
    • The fixture lives under /tmp and doesn't survive a reboot, so rebuild it from the recipe first.
    • Running the grader leaves ootd/tests/__pycache__/, which .gitignore covers.
  • Other skills have no tests. Re-read the diff against the rule it implements. For a description change, check that the trigger phrases still match how Dan asks.
  • Install checks: test install.sh without touching the real links by running CLAUDE_SKILLS_DIR=<scratch dir> ./install.sh. A skill is live when ls -l ~/.claude/skills/ shows a link into this repo and the skill appears in a session's skill list.

Public repo hygiene

• Examples use invented or generic values. Don't copy real names, sizes, measurements, addresses, IPs, hostnames or new local paths from the vault or the machine into skills, examples, commit messages or issues.
  • Two example values in ootd's templates (the ## Direction quote and the MUJI size) were reviewed on 2026-09-16 and kept on purpose. Don't replace them.
• Security details stay private. Don't publish which commands are pre-approved in other repos' settings.
• Scan before pushing. Check new content for secrets and personal details, including removed lines and commit messages: git log -p origin/main..main. History is permanent once pushed.
• Commit identity: commits use Dan's public email.