Prompt Library

Prompt Library Skill

Extract and save effective prompts from work sessions.

Process

1. Scan session for prompts that:

   • Achieved the desired outcome efficiently
   • Used clever techniques or patterns
   • Could be reused for similar tasks
   • Demonstrated good prompt engineering

2. Categorize each prompt:

   • CODE - Code generation, refactoring, debugging
   • REVIEW - Code review, architecture analysis
   • DOCS - Documentation, READMEs, comments
   • TEST - Test generation, coverage
   • DEPLOY - CI/CD, packaging, deployment
   • DESIGN - UI/UX, system design
   • DATA - Data processing, analysis
   • META - Prompts about prompting

3. Extract template - generalize specifics into placeholders

4. Pre-write format-match check (mandatory):

   • Run grep -E '^### P-' ~/projects/DecisionLog/PROMPTS.md | head -3 to confirm the target file already contains P-YYYYMMDD-### entries.
   • If the file has zero P- entries OR has entries in a different format (e.g., named templates without P-IDs), STOP. The destination has drifted. Surface to user: "PROMPTS.md format doesn't match the skill's catalog shape (P-IDs + Category field). Which file should I write to, or should we re-route the skill?"
   • This catches path-drift before silently writing into the wrong file. Discovered 2026-05-23 — the public-notes PROMPTS.md is a templates file, NOT a catalog; writing P-ID entries there would corrupt it.

5. Append to ~/projects/DecisionLog/PROMPTS.md (in the private repo AreteDriver/DecisionLog). This is the skill's design target: P-IDs + Category field + the same CODE/REVIEW/DOCS/TEST/DEPLOY/DESIGN/DATA/META vocabulary as the Categories table above. A separate file at ~/projects/notes/PROMPTS.md exists for named template scaffolds (Universal, CI/CD Diagnostic, GitHub Profile, etc.) — that one is hand-maintained, not a /prompt-library target.

Entry Format

### P-YYYYMMDD-###: [Short Title]
**Category:** [CODE|REVIEW|DOCS|TEST|DEPLOY|DESIGN|DATA|META]
**Context:** [When to use this]

**Prompt:**

[The prompt text with {{placeholders}} for variable parts]

**Why it works:**
- [Key technique 1]
- [Key technique 2]

**Example use:**
[Brief example of how it was used]

What to Extract

Good candidates:

• Prompts that got it right first try
• Novel approaches to common problems
• Prompts with good constraint specification
• Multi-step workflows that worked well
• Effective few-shot examples

Skip:

• Simple one-off requests
• Prompts that needed multiple corrections
• Context-heavy prompts that won't generalize

Prompt Patterns to Look For

1. Constraint sandwich: Context → Constraints → Task
2. Output format spec: Explicit structure requests
3. Role assignment: "Act as X with Y years experience"
4. Negative examples: "Do NOT do X"
5. Step-by-step: Breaking complex tasks into phases
6. Few-shot: Examples before the actual request