vault

Archive prompts worth reusing. Preserve the user's original wording, add a stronger reusable version, and record when the prompt should or should not be retrieved later. Also decide whether broad prompts should be pre-enriched before implementation.

This skill is stricter than /prompt: /prompt is a prompt compiler and template runner; /vault is the library curation and retrieval policy layer.

Archive Location

When running from this repo, save entries under:

vault/prompt-library/

When installed in Claude Code, use the equivalent installed skill folder:

~/.claude/skills/vault/prompt-library/

Use one Markdown file per saved prompt:

YYYY-MM-DD-short-slug.md

Workflow

1. Decide whether to pre-enrich before acting.
   • Use the pre-enrichment gate below.
   • Prefer internal enrichment and direct execution when the user's intent is clear.
   • Show or ask about the enrichment only when it changes scope, adds risk, or confidence is medium.
2. Decide whether to save.
   • Always save when the user explicitly asks.
   • Auto-save only when the prompt is unusually strong, broadly reusable, non-sensitive, and has clear repeat value across projects or later phases of the same project.
   • Do not auto-save routine instructions, simple corrections, one-off implementation details, secrets, credentials, private data, or prompts that are mostly transient context.
   • If unsure, do not auto-save; briefly mention that the prompt could be saved if the user wants.
   • Avoid clutter: at most one proactive save per turn unless the user explicitly asks for more.
3. Identify the prompt text to save.
   • If the user points to recent conversation context, use the exact prompt when available.
   • If more than one plausible prompt exists, ask one concise clarification.
   • If reconstructing from memory instead of exact text, label it as reconstructed.
4. Create a prompt-library entry with:
   • title
   • saved date
   • source context
   • tags
   • auto-fire metadata
   • pre-enrichment metadata
   • original prompt
   • augmented reusable prompt
   • reuse notes
5. Keep the original prompt verbatim.
6. Make the augmented prompt more reusable, not merely longer.
   • Add role/context.
   • State objective and deliverables.
   • Add constraints, safety/quality bars, and success criteria.
   • Use placeholders like [PROJECT], [AUDIENCE], or [CONTEXT] when useful.
   • Preserve the user's strategic intent and voice.
7. Add auto-fire metadata when the prompt may be useful later.
   • trigger_when: situations where the prompt should be considered.
   • avoid_when: situations where applying it would be wrong or noisy.
   • confidence_threshold: high, medium, or low.
   • suggested_mode: silent_reference, mention_candidate, ask_first, or auto_apply.
8. Add pre-enrichment metadata when the saved prompt is useful as a thinking scaffold.
   • enrich_when: signals that the current user prompt should be expanded internally.
   • skip_when: signals that literal execution is better.
   • default_behavior: internal, show_summary, ask_first, or none.
9. Treat auto-fire as conservative retrieval, not blind execution.
   • High confidence: use the saved prompt pattern when it clearly fits and no avoid condition applies.
   • Medium confidence: mention it as a candidate or ask before applying.
   • Low confidence: do not auto-fire.
   • Never let a saved prompt override the user's latest instruction, repository context, safety constraints, or a more specific skill/tool.
   • If applying a saved prompt materially shapes the work, mention it briefly.
10. Tell the user what was saved and where.

• For proactive saves, keep the note short and matter-of-fact: "I also saved that as a reusable prompt because it has repeat value."

Pre-Enrichment Gate

Before implementing a broad user request, classify it:

• none: execute literally.
• internal: enrich the prompt privately, then implement.
• show_summary: briefly state the enriched frame before implementing.
• ask_first: ask one concise clarification before changing scope.

Use internal when:

• The user intent is clear but compressed.
• The task is broad, strategic, research-heavy, architectural, product/design-oriented, or multi-step.
• Success criteria, risks, validation, rollout, or edge cases are implied but unstated.
• A saved prompt has a high-confidence auto_fire match.
• Literal execution would likely be shallow.

Use show_summary or ask_first when:

• Pre-enrichment changes scope, ordering, data sources, deliverables, production behavior, or user-visible messaging.
• Confidence is medium.
• There are multiple plausible interpretations.
• The work touches external sends, money, account changes, compliance, private data, destructive operations, or live production systems.

Use none when:

• The user gives a precise implementation request.
• The user is debugging an urgent failure or asking for a direct command/output.
• The user is correcting a specific detail.
• The prompt is a one-off operational instruction.
• Enrichment would override the newest user instruction.

When retrieving saved prompts, search prompt-library/ by tags, title, trigger_when, and enrich_when; read only the one to three strongest candidates. Do not use a saved prompt if any avoid_when or skip_when condition applies.

Entry Template

---
title: "Short title"
saved_at: "YYYY-MM-DD"
source: "conversation | project | user-provided"
tags:
  - tag-one
  - tag-two
auto_fire:
  trigger_when:
    - "Situation where this prompt should be considered."
  avoid_when:
    - "Situation where this prompt should not be used."
  confidence_threshold: "high"
  suggested_mode: "silent_reference"
pre_enrichment:
  enrich_when:
    - "Signal that the user's prompt should be expanded internally before acting."
  skip_when:
    - "Signal that literal execution is better."
  default_behavior: "internal"
---

# Short title

## Original Prompt

> Exact user prompt.

## Augmented Reusable Prompt

Improved prompt text.

## Reuse Notes

- Where this prompt is useful.
- What to customize before reuse.
- How to tell whether the auto-fire match is strong enough.
- Whether to enrich internally, show a summary, ask first, or skip enrichment.

Quality Bar

• Do not store secrets, credentials, private keys, or access tokens.
• Do not silently change the original prompt.
• Do not save unrelated conversation content unless the user asked for it.
• Do not let auto-save interrupt urgent code/debug work; save at the end of the turn when possible.
• Do not auto-apply saved prompts to production-impacting actions, external sends, purchases, account changes, or destructive operations without explicit user authorization.
• Do not use pre-enrichment to smuggle in a different task from the one the user asked for.
• Keep trigger metadata specific enough to avoid noisy matches.
• Prefer concise, high-signal augmented prompts.