skills-bank

When to use

• The user invokes /skills-bank, with or without a request after it.
• The user asks whether they have a skill/prompt/agent for something, or to reuse one of theirs.
• The user asks to save, create or improve a skill or prompt.
• Do NOT use for general questions that don't involve their stored library.

Inputs to collect first

• request: what they want to do. If /skills-bank is typed alone, run step 1 with no query and show the menu in step 6.

Steps

1. Find. Run bank search "<key words from the request>". If nothing matches, retry with 1–2 synonyms; if still nothing, say so and offer step 5.
2. Pick. Show at most 5 matches as id — name (type, project). Choose the best one yourself when it's obvious; ask only if two are equally good.
3. Read. bank show <id>. Note its inputs: list and uses: dependencies (bank deps <id> if any).
4. Use it.
   • prompt → collect any missing inputs in ONE question, then bank render <id> k=v k=@file … and follow the rendered text as your instructions for this task (or show it if the user only wanted the text).
   • skill → follow its body as instructions for the current task. If the user wants it permanently available, bank export-skill <id> (becomes / next session).
   • agent → treat the body as your system brief for the task.
   • tool / connector / instruction → summarise what it says and apply it.
5. Create / improve (when asked, or when nothing fits and the user agrees):
   • new: bank create-skill "<one-sentence what + when>" --project <jobeng|buildings|interior|shared>
   • improve: bank create-skill --improve <id> "<feedback>"
   • plain prompt: bank new prompt <id> then edit the file it prints and bank validate.
6. Menu (bare /skills-bank): bank stats | head -12, then offer: search · show · use · create · export · dashboard (bank dashboard).

Output format

• One line saying which entry you used: Using <id> (<type>) from the bank.
• Then the actual work product of the task — not a description of the bank.
• After creating/exporting: the exact command run and the file path it printed.

Rules

• Always go through the bank CLI; never edit bank files directly except the one bank new just created.
• Never invent an entry id — only use ids that bank search / bank list returned.
• Ask at most one clarifying question per step; batch missing inputs together.
• Don't export skills unless the user asks — keep ~/.claude/skills small.
• If bank isn't found: tell the user to run sh ~/.skills-bank/install.sh (or pip install git+https://github.com/salahibrahimplq/skills-bank) and stop.
• Respect the entry's model: hint only when the user is calling bank run; inside Claude Code, just do the work.

Example

User: /skills-bank write a rejection email for Ahmed, site engineer at Kier, after interview → bank search "rejection email" → jobeng-rejection-email (prompt, jobeng) → bank show jobeng-rejection-email → inputs: candidate_first_name, role_title, company_name, stage, reason_category, keep_on_file → asks once: "Reason, and keep on file?" → user: "stronger candidates, yes" → bank render jobeng-rejection-email candidate_first_name=Ahmed role_title="Site Engineer" company_name=Kier stage=interview reason_category=stronger-candidates keep_on_file=yes → "Using jobeng-rejection-email (prompt) from the bank." + the finished email.