Karpathy — surface uncertainty, clean orphans, goal-loop execution

ACTIVE every coding task. State assumptions. Surface ambiguity via MCQ. Clean orphaned imports. Define success criteria + loop

Karpathy — surface uncertainty, clean orphans, goal-loop

ACTIVE EVERY coding task. 3 unique disciplines. Overlapping sections cross-reffed, not restated.

1. Think Before Coding

State assumptions before acting. Surface uncertainty explicitly.

• Assumption list required before multi-file changes.
• Two+ interpretations exist? Present all in MCQ. Don’t pick silently.
• Simpler approach exists? Say so. Push back.
• Unclear? Stop. Name what’s confusing. Ask via MCQ.

2. Clean Your Own Orphans

When edit leaves dead code: remove imports, vars, functions YOU made unused. Never touch pre-existing dead code unless asked.

Cross-ref: Ponytail §ULTRA — “touch only what you must.” This rule extends it: you MUST clean your own mess.

3. Goal-Driven Execution

Transform vague ask into verifiable goal. Loop until green.

Plan format:

1. [Step] ? verify: [check]
2. [Step] ? verify: [check]

Cap at 200 words. Cross-ref: Output-minimalism §Numerical limits.

Duplicates (not restated here)

Cross-refs

• ponytail — code minimalism, surgical changes, no defensive code
• caveman — terse prose for assumption statements
• output-minimalism — answer-first, no preamble
• minimum-everything — smallest unit per task
• Upstream: karpathy/guidelines — MIT, adapted

Edit on GitHub · Back to index