Prompt Library

by The Agent Ledger — practical AI systems for solopreneurs

"Your best prompts are assets. Treat them like code."

Most people write a great prompt, use it once, and never find it again. This skill gives your agent a structured system for capturing, versioning, and retrieving your highest-leverage prompts — so every good prompt compounds instead of disappearing.

What This Skill Does

• Captures prompts with context, variables, and performance notes
• Organizes prompts by category, use case, and quality score
• Versions prompts so you can track what changed and why
• Surfaces the right prompt when you need it
• Tracks which prompts perform and which need work

Setup

Step 1: Create Your Prompt Library File

Create prompt-library.md in your workspace root (or project folder):

# Prompt Library

## Stats
- Total prompts: 0
- Last reviewed: —
- Top category: —

## Index

| ID | Name | Category | Score | Last Used | Tags |
|----|------|----------|-------|-----------|------|
| — | — | — | — | — | — |

Step 2: Choose Your Category Set

Default categories (customize freely):

Step 3: Add AGENTS.md Standing Instruction (Optional)

Add to your AGENTS.md:

## Prompt Library

When I say "save this prompt" or "add to prompt library", immediately
capture the current prompt with full context into `prompt-library.md`.
When I ask for a prompt for a task, search the library before generating
a new one — and suggest updating or versioning it after we test it.

Step 4: Initial Capture Session

In your first session, run:

"Review this conversation and extract any prompts or prompt patterns worth saving. Add them to the prompt library with initial scores and categories."

Prompt Record Format

Each saved prompt follows this structure:

### [PL-###] Prompt Name
**Category:** research | writing | analysis | planning | coding | brainstorm | review | ops | meta
**Score:** 1–10 (see scoring rubric below)
**Version:** 1.0
**Last Used:** YYYY-MM-DD
**Tags:** tag1, tag2, tag3

**Template:**

[The full prompt text here. Use {{VARIABLE}} for substitution points.]

**Variables:**
- `{{VARIABLE}}` — Description of what to substitute

**Best Used For:**
- Specific use case 1
- Specific use case 2

**Model Notes:**
- Works well with: (model types or sizes that perform best)
- Avoid with: (models that underperform on this prompt)

**Performance Notes:**
- What makes it work
- Known failure modes or edge cases

**Version History:**
- v1.0 (YYYY-MM-DD) — Initial capture

Prompt Scoring Rubric

Score each prompt 1–10 across three dimensions, then average:

Final Score = (Clarity + Reliability + Versatility) / 3

Score thresholds:

• 8–10 → Core Prompt — High-value, use confidently, version carefully
• 6–7.9 → Solid Prompt — Works well, refine over time
• 4–5.9 → Draft Prompt — Functional but needs iteration
• Below 4 → Retire — Not pulling its weight; archive or kill

Usage Patterns

1. Save a New Prompt

"Save this prompt to the library: [paste prompt]. Category: writing. Notes: works great for newsletter intros."

Agent assigns the next ID (PL-001, PL-002, etc.), fills the record template, sets initial score as "unrated," and adds to the index.

2. Find a Prompt for a Task

"I need a prompt for [task description]. Check the library first."

Agent searches by category and tags, surfaces relevant matches with their scores, and either returns the best match or drafts a new one if nothing fits.

3. Rate or Update a Prompt

"Update PL-012 — the output was shallow. Score it a 5 for reliability. Add a note: needs more explicit output format instructions."

Agent updates the record, bumps the version to v1.1, adds a version history entry, and notes the date.

4. Version a Prompt

"Version PL-007. Here's the updated version: [new prompt text]. Reason: added chain-of-thought instruction."

Agent archives the old version in version history, replaces the template with the new version, and increments version number.

5. Library Dashboard

"Give me a prompt library overview."

Agent returns:

📚 PROMPT LIBRARY — [Date]

Total: [N] prompts
Core (8+): [N] | Solid (6-8): [N] | Draft (<6): [N] | Unrated: [N]

By Category:
  research: [N]  writing: [N]  analysis: [N]  planning: [N]
  coding: [N]    brainstorm: [N]  review: [N]  ops: [N]  meta: [N]

Most Used: [Top 3 by last-used date]
Highest Scored: [Top 3 by score]
Needs Attention: [Drafts that haven't been revised in 30+ days]

Last reviewed: [date]

6. Category Deep Dive

"Show me all my writing prompts sorted by score."

Agent returns a table of all prompts in the category with ID, name, score, version, and last-used date.

7. Build a Prompt from Scratch

"Help me build a prompt for [task]. I want it to [goals]. Add it to the library when we're done."

Agent iterates with you, tests variations, and saves the final version with initial score and your notes.

8. Prompt A/B Log

"Log an A/B test: tested PL-003 vs PL-015 for competitor analysis. PL-015 won — more structured output."

Agent adds to the A/B test log section of the library file.

9. Review Stale Prompts

"Which prompts haven't been used in 60+ days? Flag ones worth retiring."

Agent scans last-used dates, flags candidates, and asks for your decision: keep, retire, or update.

10. Export Prompt Set

"Export all my research prompts as a clean list I can share."

Agent generates a shareable markdown list (IDs stripped, personal notes removed, variables labeled clearly).

Prompt Library File Structure

workspace/
├── prompt-library.md        # Main library (index + all records)
└── prompt-library/
    ├── ab-tests.md          # A/B test log (optional)
    └── retired/
        └── archive.md       # Retired prompts (for reference)

Alternatively, split by category for large libraries:

workspace/
└── prompt-library/
    ├── README.md            # Index and stats
    ├── research.md
    ├── writing.md
    ├── analysis.md
    └── ...

A/B Test Log Format

## A/B Tests

### [Date] — [Task Description]
- **Prompt A:** PL-###
- **Prompt B:** PL-###
- **Test:** [What you asked both to do]
- **Result:** [Which won and why]
- **Action:** [Updated scoring / versioned winner / retired loser]

Heartbeat Integration

Add to HEARTBEAT.md:

## Prompt Library
- If I've mentioned saving a prompt this week but the library hasn't
  been updated, remind me to capture it.
- Weekly: flag any Draft prompts (score < 6) that are over 14 days old
  with no updates — should be improved or retired.

Cron: Weekly Prompt Review (Optional)

Schedule a weekly prompt health check:

openclaw cron add \
  --name "prompt-library-review" \
  --cron "0 9 * * 1" \
  --model "anthropic/claude-sonnet-4-6" \
  --session isolated \
  --message "Review prompt-library.md. Flag: (1) prompts scored below 5 that haven't been updated in 14+ days, (2) prompts last used 60+ days ago worth retiring, (3) unrated prompts that should be scored. Deliver a brief action list." \
  --announce \
  --tz "America/Chicago"

Integrations

Customization

Custom categories: Replace or add to the default set. Common additions: client-comms, legal-review, product, seo, ads.

Team use: Shared prompt libraries work well in collaborative workspaces. Add a contributor field to records and use the export pattern to share high-value prompts.

Model-specific libraries: If you work across multiple models, add a model field to records and filter by model when searching.

Prompt chaining: Some tasks are multi-step prompt chains. Use a chain tag and document the sequence in performance notes.

Minimal setup: You don't need the full structure. Start with a single prompt-library.md and just the template. Add scoring and versioning once you have 20+ prompts.

Troubleshooting

"The library is getting too long to search" Split into per-category files and use a master index in README.md. Tell your agent to always check the index first.

"I keep forgetting to save prompts" Add the AGENTS.md standing instruction (Step 3). Also useful: end-of-session habit of asking "did we use any prompts worth saving?"

"My scores feel arbitrary" Use the A/B test log to calibrate. After 10 tests, your score intuitions become more consistent.

"Some prompts are too context-specific to generalize" That's fine — keep them. Use specific tags (project name, client name) so they're findable without polluting general searches. Note context requirements in "Best Used For."

"How do I handle prompts with lots of variables?" Document each variable clearly. If there are more than 5 variables, consider splitting into a simpler base prompt plus a specialization layer.

Privacy Note

Your prompt library may contain context-specific language, internal business logic, or sensitive framing. It is local by default. Before exporting or sharing any prompt:

• Strip business-specific context
• Remove variable examples that contain private data
• Review for inadvertent information disclosure

The export pattern (Usage Pattern #10) is designed to help you share clean versions.

The Agent Ledger publishes practical AI systems for solopreneurs and creators. Get new skills and guides at theagentledger.com → Subscribe to the newsletter for weekly AI workflow breakdowns