karpathy-loop

When to Use

Prompt bloated/slow/expensive · need consistent quality at lower cost · production/automation prompts · skill authoring — every skill should pass karpathy-loop.

STYLE (5 Rules)

1. ID+TASK=ENOUGH — identity+task. 2. MINIMAL — no 10+ item lists. 3. FORMAT=INSTRUCT — "Respond ONLY in JSON:{...}". 4. CONSTRAINTS=FORMAT — max X chars / code only. 5. IMPLICIT CoT — no "step by step"; "Reason ONLY if ambiguous".

LOOP: Write→Measure→Cut→Repeat

1. WRITE: role + task + 1-2 examples + output. Don't optimize yet. 2. MEASURE: chars/4 ≈ tokens; score correctness/conciseness/robustness. 3. CUT: remove redundant→merge→simplify; no output change→cut. 4. REPEAT: improves+tokens down→continue; drops→revert; stagnant→new tactic.

Example: Write 680/170tok→6.3 | Cut T1 480/120→7.3 | T2 340/85→7.3 | T3 260/65→5.3 ✗ → REVERT T2. Final 340ch/85tok 7.3/10.

Compression Levels

Level	Target	What to cut
T1	20-30%	filler, transitions, "step by step"
T2	30-50%	merge redundant→bullets, remove context
T3	50-70%	template structures, shortcuts, minimal identity

Budget

ID+TASK: 20-50 | +example: +100-200 | +constraints: +50-100 | OPTIMAL: 50-300

Decision — Remove?

Output changes?→keep · else concision?→remove · else clarity?→keep · else→remove

STOP

Tokens <50 + works · 3 iterations no improvement · Fits in 1 line. NEVER: sacrifice correctness for tokens, or leave edge cases uncovered.

Anti-Rationalization

Rationalization	Red Flag	Verification
"3 iteraciones bastan, corto en T2"	Stopping at T2 in THOROUGH/complex prompt	Measure T3 — revert only if score drops (e.g., 7.3→5.3), must measure first
"Tokens ya <300, no medir"	Skipping MEASURE after CUT	tokens≈chars/4 + correctness/conciseness/robustness score each loop
"Quito constraints para ahorrar"	Removing FORMAT=INSTRUCT to save tokens	Verify output still matches JSON/schema without constraint — keep if drift

Red Flags

• Doing work without checking output format → STOP, re-read skill
• Second occurrence of same rationalization → force RED zone

Verification

• Output matches skill ## Output contract + file:line citaton
• cross-ref-check.ps1 → SKILL.md OK

Refs

lean-context · skill-improver · metricas · code-review-agent · context-watchdog · execution-mode

Absorbed (2026-09-18): prompt-engineering archived — its prompt-optimization/template workflows are covered by this skill.

Reference

Worked examples, testing patterns, edge cases, anti-patterns → docs/skills/karpathy-loop/reference.md