Keep coding agents from adding unnecessary layers, duplicate state, and fallbacks that hide failure. Deliver the complete requested behavior with proportional complexity.
With the skills CLI:
npx skills add klarkxy/no-stop-gptFor manual installation, copy the entire skills/no-stop-gpt/ directory into your
agent's skill directory. Keep SKILL.md, references/, agents/, and LICENSE.txt
together. To upgrade, replace that entire installed directory with the new copy
so retired reference files do not linger. Preserve any local customizations first.
Invoke no-stop-gpt explicitly, or let a host that supports automatic discovery
select it for substantive coding work. Installation does not guarantee activation.
| Mode | Example request | Result |
|---|---|---|
| Prevent | “Implement this feature using no-stop-gpt.” | Apply the principles while completing the code change. |
| Audit | “Review this diff for overengineering.” | Read-only keep / remove / downgrade / decide findings with evidence. |
| Sweep | “Simplify this subsystem and remove unnecessary code.” | Investigate, apply justified cuts, and verify the surviving behavior. |
A request to audit or propose repository cleanup stays read-only. An explicit cleanup request authorizes changes within its scope without another approval of intermediate findings. Retiring supported behavior or compatibility needs direction unless already authorized; publication and deployment remain separate actions.
Prefer fewer concepts among designs that satisfy the same requirements. Preserve necessary security, data integrity, accessibility, compatibility, and lifecycle guarantees. Search counts and passing tests alone do not prove something is safe to remove; no justified deletion is a valid result.
Start with SKILL.md. It links to optional design examples, focused defense and lifecycle guidance, and the consolidated Sweep workflow. Pure explanation, formatting, and unrelated prose do not trigger the skill.
Compared against upstream instructions on 2026-09-08. These are design references, not dependencies or a measured ranking; the guidance here is written for this skill.
| Reference | Adopted here | Adaptation |
|---|---|---|
| Karpathy Guidelines | Observable success criteria and local conventions. | Keep useful single-use boundaries; clarify only material choices. |
| Anthropic code-simplifier | Readability and ease of debugging when simplifying control flow. | Follow the target project's language and style conventions. |
| gtapps/simplify | Concrete efficiency checks, semantic comparison, and reconciliation of overlapping findings. | No mandatory agent team; recheck stale findings and honor intentional behavior changes. |
| HERO-Anti-OverDefense | Already covered: reachable cases and checks that inform a decision. | Establish actual trust boundaries and preserve necessary protections. |
The entrypoint covers success criteria; the existing design, antipattern, and Sweep references contain the conditional examples. No additional runtime is required.
The rules are adapted from 2025–2026 community practice and classic texts. Thanks, in no particular order:
- Andrej Karpathy — Simplicity First / Surgical Changes
- William of Ockham — the razor cuts entities, not requirements; count concepts, not lines
- ML ablation studies — remove one thing, hold the rest constant, measure before you decide
- HERO-Anti-OverDefense — reachable vs constructible
- Joe Armstrong / Erlang — Let it crash (fail the isolated unit; recover outside it)
- LessWrong, pathological guardrailing — fail early and visibly
- ponytail — reuse ladder, ceiling comment, runnable check (this skill rejects its "laziness" philosophy)
- Sandi Metz — The Wrong Abstraction
- Robert C. Martin — Clean Code: abstraction levels, flag-argument tradeoffs, meaningful names, and local cleanup (used as contextual heuristics, not universal deletion rules)
- John Ousterhout — A Philosophy of Software Design
- Rob Pike — Go proverbs
- Martin Fowler — YAGNI's four costs
- Rich Hickey — Simple Made Easy
- Casey Muratori — semantic compression
- Carson Gross — Locality of Behaviour
- grug — cut-points
- TigerBeetle TIGER_STYLE & NASA's Power of 10 — where heavy defense is the optimum
- Mitchell Hashimoto — rules as failure log
- OpenClaw AGENTS.md — a fallback is a product decision
SATA 2.0 (modified), based on the original SATA license. Star the project and thank the authors—even quietly in your mind. This project's version clarifies that starring and thanking are voluntary.