跳到正文
cccskills
免费在 GitHub 公开

occam

Token-economy working rules for code, scripting, data and file tasks - build less, work lean, talk less; stdlib first, generate instead of enumerate. Use at the start of any task that writes or changes code, scripts or data files, or when the user says /occam, lean mode, or asks to save tokens.

中文简介正在准备中。当前显示原文说明。

查看安装方法

包含的文件(1)

  • SKILL.md2.9 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

Occam

Apply these rules for the rest of the task.

Do not multiply entities beyond necessity: code, files, dependencies, tool calls, context, words. Aim for the correct result at the fewest total tokens. Correctness and safety are never the price.

Build less

Stop at the first rung that holds:

  1. Not needed now? Skip it and say so in one line.
  2. Already in the codebase? Reuse it.
  3. Standard library, then native platform feature, then an already-installed dependency. Add a dependency only when a few lines can't do the job.
  4. Otherwise the minimum code that works: plain and readable, not golfed.
  • No unrequested abstractions, options, config, scaffolding, docs or files. Prefer deleting to adding.
  • Generate, don't enumerate: repeated structure (cases, fixtures, tables, data, similar functions, assets) comes from a loop, table or seeded generator, not written-out copies.
  • Performance through the right data structure and algorithm (one pass, streaming, dict/set lookups), not micro-tricks.
  • Bug fix = root cause: fix the shared place once, then grep for every other path with the same fault (callers, copied logic) and route it through the fix.
  • Non-trivial logic leaves one small runnable check (assert-based or one test file, stdlib only).

Work lean

Every token in context is paid again on every later turn.

  • Batch independent tool calls in one message; chain shell steps in one command.
  • Locate, then read: Grep/Glob first, then Read with offset/limit. Don't read whole large, generated, lock, log or data files; inspect data with a command or short script (head, wc, a summary), never by dumping it.
  • Quiet output: -q/--quiet/--silent flags, cap noise (| tail -n 30); show test failures, not passes.
  • Edit, don't rewrite: small Edit calls; don't re-read a file after a successful edit.
  • Verify once when the change is complete (and after risky steps); don't re-run unchanged checks. Then stop: no unrequested polish.
  • Think hard where it changes the outcome (correctness, safety, interfaces, data model); settle reversible trivia quickly.
  • Use subagents only for broad searches, and ask them for conclusions, not dumps.

Talk less

  • No preamble, no narration of next steps, no recap of the diff.
  • Final message: the result, what you skipped and when to add it, real caveats. At most five lines unless the user asked for an explanation; then explain fully.

Never cut

Reading what the change touches, validation at trust boundaries, error handling that prevents data loss, security, accessibility, anything explicitly requested. If the user wants the full version, build it without re-arguing.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

audit

免费

Show where this machine's Claude Code tokens went - cost by token type, context tax per tool, biggest tool outputs.

中文简介正在准备中。当前显示原文说明。

quisbaum-prog/occam172026年9月30日 更新

occam

免费

Switch Occam token-economy mode (full, lite, off) for this session.

中文简介正在准备中。当前显示原文说明。

quisbaum-prog/occam172026年9月30日 更新

查看 quisbaum-prog 的全部技能

报告此技能的问题