Agent Skills: Skill Perfection

Skill Perfection

A systematic process for auditing and fixing plugin skills in a single pass.

Core Principle

Audit + Fix in One Pass: When you find an issue, fix it immediately, then continue. This eliminates redundant research and multiple iterations.

Process Overview

[Optional Preflight] → Audit+Fix (single pass) → Verify (spot-check) → Report

Phase 1: Preflight (Optional, Advisory)

A Python-based preflight script is bundled for Python-heavy skills only.

When to Use Preflight

| Skill Content | Use Preflight? | |---------------|----------------| | Mostly Python code blocks | ✅ Yes | | Mixed Python + other languages | ⚠️ Optional (Python blocks only) | | Non-Python (JS, Rust, Go, etc.) | ❌ Skip, go to Phase 2 | | Skill about the preflight script itself | ❌ Skip (conflict of interest) |

Running Preflight

uv run python ${SKILL_DIR}/scripts/preflight.py <path-to-skill.md> --no-urls

Interpreting Results

| Result | Action | |--------|--------| | ✅ PASSED | Good signal. Proceed to Phase 2, trust syntax is valid. | | ❌ FAILED with clear errors (syntax error at line X) | Fix those specific issues, then proceed to Phase 2. | | ❌ FAILED with confusing/many errors | Ignore preflight entirely. Proceed to Phase 2, let LLM verify. | | Script crashes or hangs | Ignore preflight entirely. Proceed to Phase 2. |

Key Rule: Preflight is Advisory

The preflight script is a helper, not a gatekeeper. If it produces confusing output, skip it. The LLM-based Phase 2 is always authoritative.

Signs to ignore preflight:

• More than 10 errors on a skill that "looks fine"
• Errors that don't make sense (line numbers don't match)
• Python tracebacks from the script itself
• Timeouts or hanging

When in doubt, skip preflight and let the LLM verify everything.

Phase 2: Audit + Fix (Single Pass) - THE CORE

This is the main phase. Work through the skill file section by section.

2.1 Identify the Technology

From the skill's content, identify:

• Primary technology/framework (e.g., LangGraph, React, Axum, Go-kit)
• Version constraints if mentioned
• Official documentation domain

2.2 Research Strategy

Batch similar lookups to minimize web calls:

Example: For a skill with 10 import statements from the same package,
do ONE search that covers them all, not 10 separate searches.

Documentation priority:

1. Official docs (use site: filter)
2. Official GitHub repo (examples, tests, README)
3. Package registry (PyPI, npm, crates.io, pkg.go.dev)
4. Official blog/changelog

2.3 For Each Verifiable Item

| Item Type | What to Verify | |-----------|---------------| | Import/require statement | Package exists, path is current, not deprecated | | API call | Signature matches official docs, parameters correct | | Code example | Would execute, complete imports, correct syntax | | URL/link | Accessible (WebFetch), points to claimed content | | Version claim | Current/accurate | | Best practice claim | Aligned with official recommendations |

2.4 Fix-As-You-Go

When you find an issue:

1. STOP auditing that section
2. FIX the issue immediately (Edit tool)
3. LOG the change in memory: {location, old, new, reason, source_url}
4. CONTINUE auditing

All severities get fixed in one pass - don't defer anything.

2.5 Syntax Verification (LLM-based)

For code blocks, verify syntax by reading carefully:

Python: Check for matching parentheses, correct indentation, valid syntax JavaScript/TypeScript: Check for matching braces, valid syntax, correct imports
Rust: Check for matching braces, semicolons, valid syntax Go: Check for matching braces, correct package structure Any language: Apply your knowledge of that language's syntax rules

If unsure about syntax validity, note it but don't block on it - focus on semantic correctness against official docs.

Phase 3: Verification (Spot-Check)

After completing Phase 2:

1. Re-read the modified skill file
2. Spot-check 3-5 items you fixed (not everything)
3. Verify URLs you added/modified (WebFetch)

If spot-checks pass → Proceed to Phase 4 If spot-checks fail → Fix those specific items, re-check only those

Do NOT do a full re-audit. You already verified everything in Phase 2.

Phase 4: Generate Report

Create a concise report (<100 lines):

# Skill Perfection Report

**Skill**: {skill-name}
**Date**: {date}  
**Version**: {old} → {new}
**Status**: ✅ PERFECTED | ⚠️ NEEDS REVIEW

## Summary
- Items verified: {N}
- Issues found and fixed: {N}

## Changes Made

### High Priority
| Location | Change | Reason | Source |
|----------|--------|--------|--------|
| line 45 | `old` → `new` | why | [docs](url) |

### Medium/Low Priority
| Location | Change | Reason | Source |
|----------|--------|--------|--------|
| line 12 | `old` → `new` | why | [docs](url) |

## Verification
- [x] All imports verified against official docs
- [x] All API signatures match current documentation
- [x] Code examples are complete and correct
- [x] All URLs accessible

## Sources
1. {url}
2. {url}

Save to: {skill-directory}/PERFECTION_REPORT.md

Efficiency Guidelines

| Metric | Target | |--------|--------| | Web searches | <20 per skill (batch similar items) | | Iterations | 1 audit+fix pass + 1 spot-check | | Report length | <100 lines |

Anti-Patterns

❌ Run preflight on non-Python skills ✅ Skip preflight, let LLM verify

❌ Trust confusing preflight output
✅ Ignore preflight when output doesn't make sense

❌ Separate audit pass, then separate update pass ✅ Fix issues as you find them

❌ Full re-audit after fixes ✅ Spot-check only changed items

❌ Research same API multiple times
✅ Batch similar lookups, cache results

Bundled Scripts

scripts/preflight.py

Purpose: Quick syntax check for Python-heavy skills Limitation: Only useful for Python; advisory only

# Use for Python skills
uv run python ${SKILL_DIR}/scripts/preflight.py skill.md --no-urls

# Skip for non-Python skills or if output is confusing

If preflight causes problems, ignore it entirely. The LLM-based audit is always the authoritative verification.