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。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/TokenRollAI/llmdoc --skill prune命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
Explicit V3 bootstrap for repositories that do not already have valid llmdoc knowledge.
Default V3 operating skill for llmdoc-enabled projects. Route discovery — exploring the codebase, locating a concept or contract, judging the blast radius of a change — through the llmdoc CLI instead of broad file crawling.
Explicit V3 semantic verification and sync of existing llmdoc knowledge against the current repository.
正在显示 SKILL.md
| name | prune |
| description | Explicit V3 convergence pass that removes duplicated, fragmented, or low-value reconstructable llmdoc content. |
| argument-hint | [--scope <topic|path...>] [summary] |
Use this command only when existing llmdoc/ knowledge needs convergence after growth, duplication, fragmentation, or accumulation of reconstructable implementation inventory.
Load the llmdoc skill before broad exploration. CLI commands below run as npx -y @tokenroll/llmdoc <cmd>.
An explicit /llmdoc:prune invocation authorizes this run to:
llmdoc/llmdoc/meta.json.llmdoc-tmp/investigations/ when neededThis command does not authorize source-code edits.
git status -- llmdoc/ must be clean before the first formal write.git checkout -- llmdoc/ (plus deleting any newly created files under llmdoc/); never hand-edit files back.validate fails after pruning writes and cannot be repaired in-run, roll back the prune write-set before reporting failure.Run prune --report.
Decide the convergence plan with recorder.
Re-validate the result.
validate.prune --report and compare document/token scale with the first report.code.paths. Do not attach unrelated paths merely to preserve a coverage metric; call out any intentional coverage reduction.commit -m "<message>", which fingerprints the surviving docs and lands the meta.json follow-up commit automatically.success only when durable knowledge density or routing materially improves. Refresh convergence only when scale declines without losing justified mappings; otherwise repair, roll back, or report no_change as appropriate.prune updates convergence state only on successful validated convergence.prune must not advance the full baseline unless it explicitly performs a full successful sync as part of the same run.success: knowledge density or routing materially improved, the result validated, and convergence state was updated when applicable.no_change: the declared scope was fully verified and no justified convergence action remained.dry_run: the user asked for a dry run, or only prune --report/planning output was produced without writing llmdoc/; do not advance state.incomplete: evidence was insufficient, user input is required, or the request belongs to a different explicit workflow; roll back writes and do not advance state.failed: prune failed and writes were rolled back.Always report:
prune --report signal that justified the runvalidate and commit results