/sigil

Meta-Prompting & Self-Improvement Patterns

Purpose: load this when improving Sigil itself, not during ordinary skill generation. It captures optional self-improvement techniques for ATTUNE, validation, and future evolution.

Contents

1. Prompt optimization
2. Self-correction patterns
3. Context engineering
4. Automatic rule generation
5. Feedback loop design
6. Incremental roadmap

Prompt Optimization

DSPy-Style Optimization

Use a closed loop:

prompt -> run -> evaluate -> refine prompt -> run again

Potential Sigil uses:

• improve description routing quality
• compare template variants
• refine discovery heuristics from outcomes

TextGrad-Style Optimization

Treat review feedback as a natural-language gradient:

initial prompt -> output -> critique -> updated prompt

Use this only when repeated critique clearly improves results.

Self-Correction Patterns

Mistake Ledger

Track recurring failures in a structured log:

## Mistake Ledger

| Date | Failure Pattern | Cause | Fix | Prevention Rule |
|------|-----------------|-------|-----|-----------------|
| YYYY-MM | missing tests | VERIFY skipped | add test check | F-test-required |

Use it to avoid repeating the same generation defects.

Reflection Loop

Generate -> Self-Review -> Identify Issues -> Regenerate

Options:

Constitutional Guardrails

Keep a compact rule set for self-review:

1. Skills MUST mirror project conventions.
2. Skills MUST NOT introduce security risk.
3. Skills SHOULD stay easy to load and selective to read.

Context Engineering

Spec-First Pattern

spec -> local rules -> generation -> review -> feedback

Use this when the skill itself is complex or safety-sensitive.

Context Budget

If a generated skill requires too much inline context, split or externalize detail into reference/.

Automatic Rule Generation

Useful source flows:

1. existing code -> convention extraction -> skill or CLAUDE.md
2. CI failures -> recurring failure pattern -> preventive skill
3. PR review comments -> repeated feedback -> new project rule

Feedback Loop Design

Three levels:

1. structural quality -> automatic validation
2. semantic quality -> self-review or external review
3. practical quality -> ATTUNE over time

Map these back to Sigil:

• structural quality -> validation-rules.md
• semantic quality -> recraft / review loop
• practical quality -> skill-effectiveness.md

Incremental Roadmap

1. Add Mistake Ledger to the journal.
2. Run Self-Refine inside VERIFY for weak drafts.
3. Track weak description activation and propose rewrites.
4. Measure context cost and recommend skill splitting when needed.