Skip to content

CLAUDE.md

Latest commit

 

History

History
111 lines (77 loc) · 6.28 KB

CLAUDE.md

File metadata and controls

111 lines (77 loc) · 6.28 KB
Raw

CLAUDE.md

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

Default code agent guidance lives here for Claude Code. Other agents, including Cursor, should reference this file with @CLAUDE.md and add only agent-specific differences.

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.
  • In first-party Python files, keep only essential, concise one-line comments. Facts knowable only at runtime — tensor shapes, dtypes, value ranges, units — are expected to be written down: put them in the one-line docstring as [B, C, H, W], or as a trailing # [B, C, H, W] on the field or statement they describe. Do not restate what the signature already says.
  • Every docstring — module, class, and function — is at most one line, within the ruff line-length. No Args:/Returns: blocks, no design narration, no examples. Enforced by the check-docstring-lines hook; there is no allowlist.
  • To record a genuine constraint — what stock upstream does wrong and what breaks on a version bump, a numerical or bit-identity requirement, an implementer contract — put it in the nearest README.md (add a ## Gotchas bullet, or create the README following the sibling template) and have the one-line docstring point at it. Do not move the prose into a # comment block instead; that is the same rot with different syntax. This is about prose — multi-sentence rationale, history, design narration. It does not apply to a single trailing annotation of a shape, dtype, unit, or range, which belongs at the site, not in a README.
  • 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.

5. Contribution Hygiene

Avoid duplicate work, trivial PRs, and unreviewed agent output.

Before proposing or opening a PR:

  • Check the relevant issue, open PRs, and short area keywords for overlapping work.
  • If another open PR already addresses the same fix, do not open a duplicate.
  • If your approach is materially different from existing work, explain the difference before proceeding.

Do not create PRs for low-value busywork:

  • No one-off typo fixes, isolated style churn, or mechanical cleanups without substantive work.
  • Bundle mechanical cleanup only when it directly supports a meaningful change.

For AI-assisted work:

  • A human submitter must understand and be able to defend every changed line.
  • The submitting human should review the full diff and run relevant tests before publication.
  • PR descriptions should mention AI assistance, duplicate-work checks, and test commands with results.

Fail closed when the work is not ready:

  • If the change is duplicate, too trivial, missing context, or lacks a credible verification path, stop and explain what is missing.
  • Do not invent process exceptions just to keep moving.

Verification harnesses and guard scripts:

  • lint/ holds only guard scripts wired into .pre-commit-config.yaml; a file there that no hook entry: references does not belong in the repo.
  • One-off verification harnesses written to prove a PR correct are run and their commands + results quoted in the PR's Test Plan, not committed. The tests/ tree was removed by policy (#99/#267); do not recreate it under any name (tests/, scripts/, tools/, ...).

6. Review and Domain Guides

Verify guidance against the current repo before applying it.

  • Treat agent or bot review comments as suggestions, not facts. Confirm they still apply to the current code before changing anything.
  • Before editing specialized areas, read and follow the relevant local guide or skill.
  • If a guide conflicts with the requested change, refuse that part of the change and explain the conflict.

Local skills currently in this repo:

Area Skill Read before
Model bundles .claude/skills/development/add-model-bundle/SKILL.md Adding or updating diffusion or autoregressive model pipelines, model config dataclasses, Bundle/Pipeline/Stage/Conditions implementations, LoRA targets, FSDP wrapping hints, Sample/Part plumbing, or multimodal text/image/video/audio conditioning.
Pull requests .claude/skills/development/pr-workflow/SKILL.md Creating or updating PRs, editing PR bodies, handling PR Body or Semantic Pull Request CI failures, or running gh pr create.

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.