karpathy-guidelines
Skill1 starsChanged 16 days agoMITView on GitHub
Umbrella behavioral discipline for the most common LLM coding mistakes — overcomplication, sprawling diffs, hidden assumptions, and calling work "done" without proof. Use at the START of any non-trivial coding task to run the four pre-flight biases (think-before-coding, simplicity, surgical change, goal-driven verification) and route to the dedicated skill for each. Do NOT use for the deep drill on one discipline — go straight to that sibling — simplicity-first, small-diff-surgical-change, surface-assumptions, or verification-before-advance.
136 other projects use a version of this.
Adopted by
Most recently taken up by
- Floppa2003/skillstheir own version, first seen Nov 16, 2026
- KPISTOLAS/claude-skillstheir own version, first seen Nov 16, 2026
- Scratchydisk/claude-skillstheir own version, first seen Nov 16, 2026
- vparsatwar-git/claude-skillstheir own version, first seen Nov 16, 2026
- pwizzang/claude-skillstheir own version, first seen Nov 16, 2026
- troia-51/claude-skillstheir own version, first seen Nov 16, 2026
Used with
Projects with it also use huashu-design, github-cli and 9 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
XIYBHKSkillhuashu-design
花叔Design(Huashu-Design)——用HTML做高保真原型、交互Demo、幻灯片、动画、设计变体探索+设计方向顾问+专家评审的一体化设计能力。HTML是工具不是媒介,根据任务embody不同专家(UX设计师/动画师/幻灯片设计师/原型师),避免web design tropes。触发词:做原型、设计Demo、交互原型、HTML演示、动画Demo、设计变体、hi-fi设计、UI mockup、prototype、设计探索、做个HTML页面、做个可视化、app原型、iOS原型、移动应用mockup、导出MP4、导出GIF、60fps视频、设计风格、设计方向、设计哲学、配色方案、视觉风格、推荐风格、选个风格、做个好看的、评审、好不好看、review this design。**主干能力**:Junior Designer工作流(先给假设+reasoning+placeholder再迭代)、反AI slop清单、React+Babel最佳实践、Tweaks变体切换、Speaker Notes演示、Starter Components(幻灯片外壳/变体画布/动画引擎/设备边框)、App原型专属守则(默认从Wikimedia/Met/Unsplash取真图、每台iPhone包AppPhone状态管理器可交互、交付前跑Playwright点击测试)、Playwright验证、HTML动画→MP4/GIF视频导出(25fps基础 + 60fps插帧 + palette优化GIF + 6首场景化BGM + 自动fade)。**需求模糊时的Fallback**:设计方向顾问模式——从5流派×20种设计哲学(Pentagram信息建筑/Field.io运动诗学/Kenya Hara东方极简/Sagmeister实验先锋等)推荐3个差异化方向,展示24个预制showcase(8场景×3风格),并行生成3个视觉Demo让用户选。**交付后可选**:专家级5维度评审(哲学一致性/视觉层级/细节执行/功能性/创新性各打10分+修复清单)。
Also in 3 projects that use this
rogue-dev-studioSkillgithub-cli
GitHub CLI authentication and workflow configuration. Use when the project needs this capability or the user / team manifest asks for it. Use for specialized `github-cli` work when listed in TEAM.yaml or explicitly requested.
Also in 3 projects that use this
OpenAISkillusing-superpowers
Use when starting any conversation - establishes how to find and use skills, requiring skill invocation before ANY response including clarifying questions
Also in 9 projects that use this
GoogleSkillponytail
Forces the laziest solution that actually works, simplest, shortest, most minimal. Channels a senior dev who has seen everything: question whether the task needs to exist at all (YAGNI), reach for the standard library before custom code, native platform features before dependencies, one line before fifty. Supports intensity levels: lite, full (default), ultra. Use on ANY coding task: writing, adding, refactoring, fixing, reviewing, or designing code, and choosing libraries or dependencies. Also use whenever the user says "ponytail", "be lazy", "lazy mode", "simplest solution", "minimal solution", "yagni", "do less", or "shortest path", or complains about over-engineering, bloat, boilerplate, or unnecessary dependencies. Do NOT use for non-coding requests (general knowledge, prose, translation, summaries, recipes).
Also in 9 projects that use this
LLl0k0laDSkillidea-refine
Refine ideas through structured divergent and convergent thinking. Use "idea-refine" or "ideate" to trigger.
Also in 4 projects that use this
OpenAISkillexecuting-plans
Use when you have a written implementation plan to execute in a separate session with review checkpoints
Also in 12 projects that use this
- +3 more from openai
Other versions136 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 136
- forrestchang/andrej-karpathy-skills★ 217ksame name
- multica-ai/andrej-karpathy-skills★ 214ksame name
- ghostfolio/ghostfolio★ 9.7ksame name
- OpenBMB/PilotDeck★ 5ksame name
- didi/mpx★ 4ksame name
- hzm0321/real-time-fund★ 1.8ksame name
- x-tools-author/x-tools★ 1.1ksame name
- coco-research/coco★ 797same name
- mxyhi/ok-skills★ 511same name
- vtroisWhite/andrej-karpathy-skills★ 426same name
126 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository15 other files it gives its agents.
- codebase-mapping-forgeSkill
- conformance-lintSkill
- skill-editorSkill
- spec-style-forgeSkill
- honesty-vocabularySkill
- reuse-before-createSkill
- simplicity-firstSkill
- small-diff-surgical-changeSkill
- surface-assumptionsSkill
- trace-to-consumerSkill
- verification-before-advanceSkill
- adversarial-review-loopSkill
- blind-evalSkill
- sandbox-isolationSkill
- skill-improvement-loopSkill
The file
We split this skill into the rules it gives an agent
3 rules. 0 are common, 2 are distinctive, and 1 is very specific to this project.
0
Common
Most of the 131 edited versions we read keep these.
2
Distinctive
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 24
When you feel the pull to write a lot of code fast — that is exactly when the four biases pay off.
- Very specific: line 29
Think before coding — surface assumptions: Don't assume, don't hide confusion, present tradeoffs, and never take the spec's framing about …
- Distinctive: line 43
Before writing code, run this in order:
“Common” is measured: kept by at least 6 in 10 of the 131 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Karpathy Guidelines
Behavioral guidelines to reduce the common coding mistakes an LLM makes regardless of stack, derived from Andrej Karpathy's public observations on where LLM coding goes wrong. This is the front door of the GUARD pack: it names the four failure modes and hands you to the dedicated skill for each. Read it first on any task that is more than a trivial edit.
Tradeoff: these guidelines bias toward caution over speed. For a truly trivial change, use judgment — but "trivial" is a claim you should be able to defend, not a default.
When to use
- At the start of any non-trivial implementation, refactor, or bug-fix.
- When you feel the pull to write a lot of code fast — that is exactly when the four biases pay off.
- As a checklist before you claim a task is finished.
The four biases (and where each lives)
- Think before coding — surface assumptions. Don't assume, don't hide confusion, present
tradeoffs, and never take the spec's framing about the codebase's current state on faith.
→
surface-assumptions - Simplicity first. The minimum code that solves the actual problem; nothing speculative.
→
simplicity-first - Surgical changes. Touch only what the request requires; every changed line traces to the ask.
→
small-diff-surgical-change - Goal-driven execution — verify before you advance. Turn the task into a verifiable goal, loop
until it is proven, and close with a word your evidence supports.
→
verification-before-advance+honesty-vocabulary
Steps — the pre-flight
Before writing code, run this in order:
1. Restate the goal + assumptions → verify current state (grep), name open interpretations.
2. Pick the SIMPLEST approach that meets the goal; push back if the spec over-asks.
3. Define success criteria: "goal → verify: <concrete check>".
4. Implement the narrowest diff that satisfies (3).
5. Climb the verification ladder; close with BUILD-/SMOKE-/UNVERIFIED — never "tested" unless tests ran.For a multi-step task, write the plan as a list of step → verify: check lines so you can loop
independently. Strong criteria let you proceed without asking; weak criteria ("make it work") force
constant clarification and hide failures.
Key points
- The four biases compound: a simple design (2) produces a small diff (3), which is easy to verify (4), and rests on assumptions you already surfaced (1). Skipping one usually costs you in another.
- Each dedicated skill is self-contained and goes far deeper (real search drills, worked misses, anti-patterns). This umbrella exists to make sure you reach for all four, not to replace them.
- Two more pack skills sit alongside these four:
reuse-before-create(don't reinvent what exists) andtrace-to-consumer(prove your output reaches its reader). Both are part of "think before coding."
Worked example
Task: "Add retry to the payment webhook handler."
- Think first (1): grep the handler — does a retry helper already exist? What does "retry" mean here — transport retries, or replaying a failed business action? Two readings; name them or ask.
- Reuse (side-check): the
billingmodule already has aRetryPolicyused by three callers — reuse it rather than writing a new loop (reuse-before-create). - Simple (2): wrap the existing call in the existing policy. No new
RetryStrategyinterface, no config for backoff curves nobody asked for. - Surgical (3): the diff is the wrapping call plus one import — not a reformat of the handler.
- Verify (4): typecheck ✓, build ✓, drive one failing-then-succeeding call ✓ → SMOKE-VERIFIED, not "tested."
The overeager version invents a RetryStrategyFactory, reformats the file, and reports "done and
tested" after only a compile. Each bias catches one of those.
Anti-patterns
- Diving into code before restating the goal and checking the codebase's current state.
- Adding abstraction, config, or error paths nobody asked for.
- A large, multi-concern diff for a small requirement.
- Closing with "done" / "tested" on the strength of a compile.
- Treating this umbrella as the whole discipline — the depth is in the four dedicated skills.
Related skills
surface-assumptions— bias 1, in depth.simplicity-first— bias 2, in depth.small-diff-surgical-change— bias 3, in depth.verification-before-advance— bias 4, in depth.honesty-vocabulary— the words you are allowed to close with.reuse-before-create/trace-to-consumer— the two search disciplines that pair with "think before coding."
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

