跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
设计与 UI

benchmark-praxis

Benchmark Praxis methodology against external frameworks, tools, or approaches. Use when: benchmark, compare praxis, how does praxis compare, gap analysis, benchmark vs, assess against, track changelog, watch features, inventory features, full feature sweep. Modes: full (default), quick, gap, track, inventory.

digital-stoic-org · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add digital-stoic-org/agent-skills --skill benchmark-praxis

技能说明


name: benchmark-praxis description: "Benchmark Praxis methodology against external frameworks, tools, or approaches. Use when: benchmark, compare praxis, how does praxis compare, gap analysis, benchmark vs, assess against, track changelog, watch features, inventory features, full feature sweep. Modes: full (default), quick, gap, track, inventory." allowed-tools: WebSearch, WebFetch, AskUserQuestion, Read, Write, Edit, Glob, Grep, Bash model: opus context: main argument-hint: "[quick|gap|track|inventory] <url, repo name, or framework name>" cynefin-domain: complicated cynefin-verb: analyze

Benchmark Praxis

Structured benchmarking of Praxis methodology against external frameworks/tools. 5 modes: full, quick, gap, track, inventory.

Benchmarking: $ARGUMENTS

Mode Detection

Argument starts withModeSetDepth
quick ...QuickSet 1 (4-5 dims)Lightweight, no deep research
gap ... or gap (no arg)GapSet 2 (4 dims)Self-assessment, inward-looking
track ...TrackFeature deltaChangelog diff vs features/{subject}.yaml, triage new entries
inventory ...InventoryFull feature sweepCrawl docs tree, append every missing feature as new for batch triage
anything elseFullSet 1 (all 8 dims)Deep research + verdict + actions

When to use inventory vs track: Inventory is a one-shot baseline build (or rare refresh). Track is recurring delta maintenance against an existing baseline. Run inventory first for any new subject; track keeps it fresh afterward.

AskUserQuestion Guard

After EVERY AskUserQuestion call, check if answers are empty/blank. If empty: output "Questions didn't display (known bug)" and present options as numbered text list instead.

Workflow

0.Scope → 1.Research → 1b.Feature delta (track only) → 1c.Feature sweep (inventory only) → 2.Score → 3.Verdict → 4.Actions → 5.Persist

0. Scope

  • Source: URL, repo, article, or framework name from $ARGUMENTS
  • Goal emphasis: AskUserQuestion — which matters most? Learning / Positioning / Import / Gap / Security
  • Prior art: Read $PRAXIS_DIR/thinking/benchmarks/inventory.yaml — has this been benchmarked before? If yes, reference it, focus on what's new/changed
  • Track mode: also Read $PRAXIS_DIR/thinking/benchmarks/features/{subject}.yaml if present — this is the durable feature-level state for the subject
  • Inventory mode: Read features/{subject}.yaml if present (will be diffed against full doc sweep). If absent, this is a first-time baseline — start fresh.

Skip AskUserQuestion if $ARGUMENTS already covers goal.

1. Research (skip for quick mode)

  • WebSearch + WebFetch: Source material, docs, README, design decisions
  • Codebase read: If Claude Code plugin/framework, read its CLAUDE.md, skill architecture
  • Output: "What It Is" summary (2-5 sentences)

Quick mode: Skip deep research, use what's available from $ARGUMENTS + quick web check.

1b. Feature delta (track mode only)

Goal: keep a per-subject feature ledger fresh and triage new entries into adopt/cherry-pick/watch/reject.

  1. Fetch changelog/release notes via WebFetch from canonical source (e.g., CC docs changelog, GitHub releases). Default window: since last_reviewed in features/{subject}.yaml, or last 8 weeks if first run.
  2. Extract candidate features: each release entry → one feature row. Capture name, version-first-seen, one-line description, source URL.
  3. Diff against existing ledger: compare candidate names against features/{subject}.yaml entries. Three buckets:
    • New: not in ledger → append with status new
    • Changed: in ledger, but description/version updated → update fields, set status to review if it was watching
    • Unchanged: skip
  4. Triage new entries using the rubric in reference.md (Feature Triage Rubric). Each new feature gets a status: adopt, cherry-pick, watch, reject, or new (if undecided — flag for human review). Include why (one line).
  5. Update ledger: write back to features/{subject}.yaml with new entries + bumped last_reviewed date.

Skip steps 2-5 if the changelog has no entries since last_reviewed.

1c. Feature sweep (inventory mode only)

Goal: build (or refresh) a complete feature baseline by crawling the full documentation tree, not just the changelog. Inventory mode answers "what features exist at all", track mode answers "what changed since last week".

  1. Identify the docs tree root for the subject (e.g., Claude Code: https://docs.claude.com/en/docs/claude-code/). Ask user to confirm the root URL if not obvious from $ARGUMENTS. Source can be a docs site, GitHub Pages, GitHub repo docs/ folder, or wiki — all work as long as content is markdown or HTML.
  2. Enumerate doc sections to crawl:
    • Claude Code shortcut: canonical surface is skills, hooks, plugins, settings, agents, mcp, sandbox, teams, schedule, remote-control, lsp, worktree, slash-commands, permissions, cli-reference, environment-variables.
    • GitHub repo source: if root is a GitHub repo, prefer gh api repos/{owner}/{repo}/git/trees/{branch}?recursive=1 and filter to paths matching docs/**/*.md or **/README.md. Fallback: gh api repos/{owner}/{repo}/contents/{docs-path} to list a single directory. Each .md file = one section.
    • Docs site source: WebFetch the index/landing page first, extract the navigation/sidebar links, propose the section list to the user.
    • Other: ask the user to enumerate sections manually.
  3. Crawl each section:
    • GitHub markdown: WebFetch via https://raw.githubusercontent.com/{owner}/{repo}/{branch}/{path} for clean unrendered markdown (avoids JS/HTML noise).
    • Docs site: WebFetch the rendered HTML page directly.
    • Per-fetch prompt: "List every distinct feature, setting, env var, hook event, CLI flag, or capab
阅读完整来源 ↗