菜单
Engineering guardrails for every change — the behavioral base of the policy spine. Make no unrequested assumptions, no over-engineering, no orthogonal edits, and match the surrounding code. Apply on every edit, plan, and review.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Measure cost and context before justifying any harness change (P7). Read /context, /cost, /status and the tool-failure log to get before/after numbers; roll back anything that worsens a metric. Use when adding tools, evaluating a phase, or diagnosing quota burn.
GitNexus is a conditional accelerator, not a default navigation tool. Use it early only when the task already signals large-repo, cross-module, blast-radius, or unclear-flow uncertainty. Every candidate must be verified with rg/Probe before reading deeply or editing.
Offload code search to deterministic tools instead of hand-grepping or reading whole files (P1). Climb the ladder lexical → structural → semantic, stopping at the first rung that answers. Use whenever locating symbols, call sites, definitions, or concepts in a repo.
Compress output to the minimum that conveys the result (P8). Drop preamble, prose, and whole-file re-prints; deliver code, diffs, and file:line refs. Use on every response, scaling density to task size (lite / full / ultra).
Offload mechanical and type-aware edits to deterministic tools instead of LLM file-rewrites (P2). Use ast-grep --rewrite for repeated/mechanical transforms and Serena replace_symbol for single type-aware symbol edits. Apply whenever a change is rule-shaped rather than judgment-shaped.
Portable code-review checklist. Use before a commit/PR or after a non-trivial change. Run it in a separate read-only helper when your runtime supports one; Sober ships the checklist, not a fixed reviewer.
| name | karpathy |
| description | Engineering guardrails for every change — the behavioral base of the policy spine. Make no unrequested assumptions, no over-engineering, no orthogonal edits, and match the surrounding code. Apply on every edit, plan, and review. |
The behavioral foundation under P0–P8 (AGENTS.md). These four catch the failures that quietly burn the most quota: invented requirements, 50→500-line over-builds, unrelated churn, and code that fights the codebase's style.
Ecosystem adoption shows the biggest win is a behavioral contract, not a tool. These guardrails make every LLM action cheap to verify and cheap to roll back (the invariant): a minimal, on-style, scope-true diff is trivial to review and revert. Everything in P0–P8 builds on this.
