hk-ipo/AGENTS.md

AGENTS.md

Foundation Guidelines

Source: https://github.com/forrestchang/andrej-karpathy-skills/blob/main/skills/karpathy-guidelines/SKILL.md License: MIT

Behavioral guidelines to reduce common LLM coding mistakes, derived from Andrej Karpathy's observations on LLM coding pitfalls.

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.

Function Extraction

Exchange Mesh code should avoid over-splitting tiny functions. Prefer functions that carry real business meaning, encode a reusable exchange rule, remove meaningful duplication, or isolate behavior that deserves direct testing.

Keep one-off field reads, single-call wrappers, simple fallback selection, and local default/override logic inline at the point of use. Avoid helpers whose body is effectively a field access, a parser call, or a call to another helper; they make readers jump around without improving the design.

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]

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

Git Workflow (Automatic Commits)

• After each completed repository change, create a focused git commit automatically before the final response.
• Use small, focused commits.
• Use rich multi-line commit messages so git log is the primary step-by-step history for this repo.
• All generated code and documentation outputs should be written in English by default.
• Commit messages should use:
  • a short imperative subject line
  • a blank line
  • concise body sections such as Request:, Changes:, Verification:, and Next useful context: when relevant
• Do not wait for the user to ask for a commit.
• Before committing, run a relevant verification command when practical.
• Include all files directly related to the completed project change, including project-local .codex skills, schema, scripts, data snapshots, memos, and documentation.
• Do not include unrelated dirty files in a commit.

Verification

• Do not run go test ./... by default in this repo.
• Many Go tests here are live API, websocket, long-running, or code-generation checks and may rewrite tracked files or stay connected for a long time.
• Default Go verification should be compile-only:
  • Prefer go test ./... -run '^$' to compile packages and test files without executing test functions.
  • When the change is isolated, prefer a narrower compile-only command such as go test ./path/to/pkg/... -run '^$'.
• Only run actual test functions when one of these is true:
  • the user explicitly asks for it
  • the specific tests are known to be short, local, deterministic, and side-effect-free
• If actual test execution is skipped because of these rules, say so in the final response and report the compile-only verification that was used instead.

Logging

• Use log.Capture only for actual errors or actionable abnormal failures. It is routed to Sentry and will appear in the error list.
• For expected state changes, successful fallbacks, routine diagnostics, and non-error informational events, use log.Debug or another non-Sentry logging path instead.

Context Recovery

• Do not maintain a separate session-notes or handoff-log file.
• Use git log as the persistent project history.
• At the start of a new thread or whenever prior context matters, inspect recent commit subjects and commit bodies before making assumptions.
• Prefer checking both the current worktree and recent git history so uncommitted local state and committed project history are both visible.