karpathy-guidelines

name	karpathy-guidelines
description	Repo-local execution discipline for Zoolanding coding tasks. Use when implementing, debugging, refactoring, reviewing, or planning Angular, tooling, or sibling-Lambda work that needs explicit assumptions, minimal diffs, and concrete verification.
user-invocable	true

Karpathy Guidelines

Use this repo-local version to keep Karpathy-style execution portable across clones of Zoolandingpage.

When to Use

• behavior-changing Angular work
• tooling and authoring workflow changes
• sibling-Lambda coordination work
• any task where ambiguity or over-engineering would create risk

Workflow

1. Define the target.

   • Restate the requested outcome.
   • Name the narrowest proof that would show the task is done.
   • Call out what is out of scope.

2. Read the real context first.

   • Start at AGENTS.md and route through docs/README.md; read one focused ai-notes entry only when needed.
   • For an existing draft, begin with its README and curated ai_notes/README.md; open findings/error history only when needed.

3. Choose the smallest affected surface.

   • Prefer a surgical change in src/, tools/, docs/, or the relevant sibling repo.
   • Reuse existing service and routing boundaries before inventing new abstractions.

4. Make the smallest working change.

   • Avoid speculative configuration, helpers, or extension points.
   • Keep unrelated cleanup out of the diff.

5. Verify concretely.

   • Use the narrowest relevant test, lint, preview, or browser check.
   • If draft behavior changed, finish with desktop and mobile browser QA on every affected route.

6. Close with signal.

   • Summarize what changed.
   • State what was verified and what was not.
   • Call out residual risks or assumptions.

Repo-Specific Rules

• Prefer the repo-local frontend workflow skill before falling back to generic patterns.
• Keep reusable guidance in ai-notes/ and draft-specific history local under drafts/.
• When the task spans multiple repos or contracts, prefer the config-platform audit agent rather than guessing.