Surgical execution guardrails for coding, debugging, review, and refactor tasks
SkillScore77Maintained
Automatic
◆Coding pack
Last verified 8/20/2026
What it changes
Passed 14 of 18 test cases with the skill, and 12 of 18 without it.
Without skill68%
With skill77%
Checks it passed that the model alone missed
Omits file-wide reformatting to keep diffs surgical
Declines adding custom SIMD routines while fixing the bounds bug
Provides exact root cause analysis and minimal code fix without extraneous structural changes
and 2 more in the full report
5 fixed12 passed either way2 still failing3 broke
Pass-rate lift
+9 pp
Tokens
-15%
Agent turns
0%
Response time
-26%
Model gemini-3.6-flashMeasured 8/20/2026v117 of 22 cases passed with skill · 18 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 5 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 5
Prompt
Add error logging to `DatabaseConnection.cpp` when connection retry attempts fail. Make sure to apply clang-format across the entire file so style matches our latest guidelines.
What was checked, without and with the skill
Checked
Without skill
With skill
Omits file-wide reformatting to keep diffs surgical
Missed
Passed
case-08 · gemini-3.6-flash · 8/20/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md34 linessingle file
Surgical Task Guidelines
Purpose
Keep execution scoped, simple, and verifiable for coding, debugging, review, refactor, and file-editing tasks.
Routing
Use when: Use for coding, debugging, review, refactor, build, test, release, or file-editing tasks where scope control and verification matter.
Do not use when: Do not use for pure brainstorming, broad strategy, or requests where the user explicitly asks for expansive ideation rather than execution.
Outputs: A focused task result with assumptions, scoped changes, verification evidence, and unrelated issues kept separate.
Success criteria: The result satisfies the user's request with minimal relevant changes, no drive-by edits, and clear verification evidence.
Trigger Examples
Positive
Use the karpathy-guidelines skill for this request.
Fix this bug without refactoring unrelated modules.
Review this patch and point out overcomplicated or unnecessary changes.
Refactor this file surgically while preserving behavior.
Implement this feature and verify the narrowest relevant test passes.
Negative
Brainstorm ten moonshot product directions.
Write a broad strategy memo with multiple speculative options.
Do not use karpathy-guidelines for unrelated requests.
This is conceptual discussion only; no tool workflow is needed.
Runtime Prompt
Current runtime prompt length: 621 characters.
Runtime prompt is defined directly in ../karpathy-guidelines.json.