richfrem/optimize-agent-instructions

<example> <commentary>User wants their agent instruction files to follow best practices.</commentary> user: "Optimize my CLAUDE.md with Karpathy principles" assistant: [triggers optimize-agent-instructions, reads files, audits against checklist, rewrites] </example> <example> <commentary>User has stale auto-generated content in their instruction files.</commentary> user: "My GEMINI.md has a bunch of stuff that doesn't belong — can you clean it up?" assistant: [triggers optimize-agent-instructions, identifies foreign content, strips and rewrites] </example> <example> <commentary>Negative — user wants to update a specific skill, not instruction files.</commentary> user: "Improve the trigger description for my link-checker skill" assistant: [triggers os-improvement-loop, not optimize-agent-instructions] </example>

optimize-agent-instructions

Audits and rewrites the AI agent instruction files in a repo. Works on any project — not just agent-plugins-skills. The goal is files that are authoritative, concise, and guide AI behavior through explicit principles rather than hoping for defaults.

---

Phase 1 — Discovery

Run these checks silently before asking anything:

1. Which instruction files exist?

ls CLAUDE.md GEMINI.md .github/copilot-instructions.md 2>/dev/null

2. Ask the user:

• What platform(s) they use (Claude Code, Copilot, Gemini CLI) — determines which files to touch
• Any project-specific rules to preserve (coding standards, ADRs, naming conventions)

If instruction files are missing for active platforms, offer to create them.

---

Phase 2 — Audit Each File

Read each file, then score it against the Quality Checklist:

Quality Checklist

Structure

• [ ] Merges Karpathy's behavioral guidelines with project-specific instructions as needed
• [ ] No stale AI-session artifacts ("Would you like X configured?", "Here is a summary of...")
• [ ] No foreign rules from other projects (spec-kitty constitution, unrelated frameworks)
• [ ] No self-referential framing ("this file reproduces CLAUDE.md...", "this is a copy of...")
• [ ] All paths and commands are current and correct
• [ ] Platform-specific notes are present where relevant (Windows caveats, tool name mapping)

Karpathy Principles — verify all four are present:

• [ ] 1. Think Before Coding: States what to clarify before starting, what to ask. Surfacing tradeoffs.
• [ ] 2. Simplicity First: Minimum code that solves the problem. Nothing speculative.
• [ ] 3. Surgical Changes: Touch only what you must. Clean up only your own mess. Match existing style.
• [ ] 4. Goal-Driven Execution: Define success criteria. Loop until verified. Verify goals.

Platform-specific (if applicable)

• [ ] Gemini: has tool name mapping table (Read→read_file, Bash→run_shell_command, etc.)
• [ ] Copilot: is authoritative (not framed as "a copy of CLAUDE.md")

Report the audit score before rewriting. Example:

CLAUDE.md: 6/8 checks pass
  ✗ No Karpathy principles section
  ✗ Stale artifact at EOF
  ✓ No self-referential framing
  ...

---

Phase 3 — Rewrite Plan

For each file that scored poorly, propose changes:

• State what will be removed (foreign content, stale artifacts)
• State what will be added (Karpathy section, platform notes)
• State what will be preserved (valid project-specific rules)

Get confirmation before writing. Show the full proposed content for each file.

---

Phase 4 — Write

Before rewriting any files, read the Karpathy principles example at references/sample-claude-md. This file contains the authoritative representation of the principles derived from forrestchang/andrej-karpathy-skills.

Write each file using the canonical structure below.

Canonical Structure

# <Title (e.g., CLAUDE.md or Copilot Instructions)>

Behavioral guidelines to reduce common LLM coding mistakes. Merge with project-specific instructions as needed.

**Tradeoff:** These guidelines bias toward caution over speed. For trivial tasks, use judgment.

## 1. Think Before Coding

**Don't assume. Don't hide confusion. Surface tradeoffs.**

Before implementing:
- State your assumptions explicitly. If uncertain, ask.
- If multiple interpretations exist, present them - don't pick silently.
- If a simpler approach exists, say so. Push back when warranted.
- If something is unclear, stop. Name what's confusing. Ask.

## 2. Simplicity First

**Minimum code that solves the problem. Nothing speculative.**

- No features beyond what was asked.
- No abstractions for single-use code.
- No "flexibility" or "configurability" that wasn't requested.
- No error handling for impossible scenarios.
- If you write 200 lines and it could be 50, rewrite it.

Ask yourself: "Would a senior engineer say this is overcomplicated?" If yes, simplify.

## 3. Surgical Changes

**Touch only what you must. Clean up only your own mess.**

When editing existing code:
- Don't "improve" adjacent code, comments, or formatting.
- Don't refactor things that aren't broken.
- Match existing style, even if you'd do it differently.
- If you notice unrelated dead code, mention it - don't delete it.

When your changes create orphans:
- Remove imports/variables/functions that YOUR changes made unused.
- Don't remove pre-existing dead code unless asked.

The test: Every changed line should trace directly to the user's request.

## 4. Goal-Driven Execution

**Define success criteria. Loop until verified.**

Transform tasks into verifiable goals:
- "Add validation" → "Write tests for invalid inputs, then make them pass"
- "Fix the bug" → "Write a test that reproduces it, then make it pass"
- "Refactor X" → "Ensure tests pass before and after"

For multi-step tasks, state a brief plan:
` ` `
1. [Step] → verify: [check]
2. [Step] → verify: [check]
3. [Step] → verify: [check]
` ` `
(Note: Do not escape backticks in actual file, use regular markdown codeblock formatting)

Strong success criteria let you loop independently. Weak criteria ("make it work") require constant clarification.

---

**These guidelines are working if:** fewer unnecessary changes in diffs, fewer rewrites due to overcomplication, and clarifying questions come before implementation rather than after mistakes.

---

## Project-Specific Rules
<Project-specific rules. Keep existing rules from previous versions, or omit section if none exist.>

Platform-Specific Sections

For GEMINI.md only — append after the main content:

## Gemini CLI Tool Mapping

| Claude Code | Gemini CLI equivalent |
|:------------|:----------------------|
| `Read`      | `read_file`           |
| `Write`     | `write_file`          |
| `Edit`      | `replace_in_file`     |
| `Bash`      | `run_shell_command`   |
| `Glob`      | `glob`                |
| `Grep`      | `grep`                |

Skills in `.agents/skills/` use Claude Code tool names in their SKILL.md files.
When executing skills via Gemini, translate tool references using the table above.

For .github/copilot-instructions.md — title line should be authoritative:

# Copilot Instructions for <repo-name>

> Authoritative rules for all AI agents (Claude Code, Copilot, Gemini) working in this repo.
> Mirrors CLAUDE.md — keep in sync.

---

Phase 5 — Verify

After writing, re-run the Quality Checklist mentally. Confirm:

1. No stale artifacts remain
2. All four Karpathy principles are present exactly as prescribed
3. Platform-specific sections are correct
4. Files are authoritative — none frame themselves as copies of another

Report:

=== optimize-agent-instructions Complete ===

Files updated:
  ✓ CLAUDE.md       — 8/8 checks pass
  ✓ GEMINI.md       — 8/8 checks pass
  ✓ .github/copilot-instructions.md — 8/8 checks pass

Karpathy principles: ✓ all four present in all files
Stale artifacts removed: 2
Foreign content removed: 1
Platform sections added: Gemini tool mapping

---

Rules

• Never remove valid project-specific rules — move them to ## Project-Specific Rules.
• Never create files that don't exist (CLAUDE.md, GEMINI.md, copilot-instructions.md) unless the user explicitly asks
• Always read before writing — never write a file you haven't read this session
• Keep all three files in sync on shared content (Karpathy principles, Project-specific rules)
• Gemini tool mapping only goes in GEMINI.md — not in CLAUDE.md or copilot-instructions.md

---

Attribution

The four behavioral principles in this skill are derived from Andrej Karpathy's observations on LLM coding pitfalls, as distilled by forrestchang/andrej-karpathy-skills.