From readme-skill
Generates a shareable, anonymized AI-Native developer README quantifying Claude Code + Codex CLI usage depth, collaboration style, project distribution, and GitHub commit correlation. All processing is local, read-only, and anonymous by default.
How this skill is triggered — by the user, by Claude, or both
Slash command
/readme-skill:readme-skillThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
You (the AI agent invoking this skill) will read local Claude Code + Codex data,
You (the AI agent invoking this skill) will read local Claude Code + Codex data,
compute a fixed set of dimensions, and render a Markdown profile under
./output/ in the user's requested language (Chinese by default; English when
the user asks in English or explicitly requests English). You do all of the
work — read the files with Read, query sqlite via Bash, and synthesize the
prose yourself. Do not write helper scripts; the skill is the recipe.
默认行为:对外分享版 —— 项目名匿名、敏感信息脱敏。 如果用户明确说"私人版 / 不要脱敏 / show real names",跳过匿名步骤。
cd <repo-with-this-skill> # e.g. ~/Projects/Readme.skill
mkdir -p output
DATE=$(date +%Y%m%d)
Decide anonymization mode (default = on). Build an in-memory mapping
real_path → "项目 A/B/C" as you encounter project paths in later steps.
Use the same mapping consistently across all sections.
~/.claude/ + 项目 .claude/)Read ~/.claude/stats-cache.json. Extract:
| 字段 | 含义 |
|---|---|
totalSessions | session 总数 |
totalMessages | 消息总数 |
firstSessionDate | 首个 session ISO 时间 |
longestSession.{duration,messageCount,timestamp} | 最长 session |
hourCounts | {hour: count} 24h 热力 |
modelUsage[model].{inputTokens,outputTokens,cacheReadInputTokens,cacheCreationInputTokens} | 每模型 token 细分 |
dailyActivity[].{date,messageCount,sessionCount,toolCallCount} | 每日活跃 |
dailyModelTokens[].{date,tokensByModel} | 每日按模型 token |
派生量(你来算):
claude_tokens_spent = Σ (inputTokens + outputTokens + cacheCreationInputTokens) —— 真实新付费 tokenclaude_cache_read = Σ cacheReadInputTokens —— 缓存复用,反映 prompt-caching 熟练度cache_to_spent_ratio = claude_cache_read / claude_tokens_spent —— 比值越大越熟~/.claude/history.jsonl —— 每行 {display, timestamp, project, sessionId}。
# Top 15 slash commands
jq -r 'select(.display | startswith("/")) | (.display | split(" ")[0])' \
~/.claude/history.jsonl | sort | uniq -c | sort -rn | head -15
# 总条数 vs 命令条数 vs 直接 prompt 条数
total=$(wc -l < ~/.claude/history.jsonl)
cmd=$(jq -r 'select(.display | startswith("/")) | .display' ~/.claude/history.jsonl | wc -l)
echo "total=$total cmd=$cmd plain=$((total - cmd))"
记录:/effort、/plan、/skill*、/usage、/clear、/resume、/compact、/init 各自次数。
~/.claude/projects/)Each subdir is one project; per-project *.jsonl files = sessions.
The dir name encodes the absolute path with / → - (ambiguous when the
original path itself contains -).
# Top 15 by session-file count
for d in ~/.claude/projects/*/; do
n=$(ls "$d"*.jsonl 2>/dev/null | wc -l | tr -d ' ')
echo "$n $(basename "$d")"
done | sort -rn | head -15
To recover the canonical real path (so you can run git log later), read
the cwd field from the first JSONL in each dir:
head -1 ~/.claude/projects/<encoded>/*.jsonl 2>/dev/null \
| jq -r 'select(.cwd) | .cwd' | head -1
Claude Code 的 plan 文件目录不是固定值。默认在 ~/.claude/plans,
但用户可以通过 plansDirectory 改到项目工作目录下,例如
"./.claude/plans"。统计 plans 时必须先解析候选 plan 目录,不能只枚举
~/.claude/plans/*.md。
解析规则:
~/.claude/projects/*/*.jsonl 的 cwd 字段恢复 Claude Code 访问过的项目根目录。.claude/settings.local.json > .claude/settings.json > ~/.claude/settings.json > default。plansDirectory:
~/... 展开为 $HOME/...;./... 或其他相对路径按该项目根目录解析。~/.claude/plans。*.md 真实路径去重后,再统计 plan 数量和标题。# Plan titles (first # heading of each plan) from all resolved plan dirs.
# Include ~/.claude/plans plus any per-project plansDirectory targets.
# Count plan files by file count, not by title extraction success.
plan_count=<resolved-plan-file-count>
for f in <resolved-plan-files>; do
awk '/^# / { sub(/^# /, ""); print; exit }' "$f"
done
ls ~/.claude/skills/ | wc -l # skills installed / authored
ls ~/.claude/tasks/ | wc -l # tasks tracked
ls ~/.claude/todos/ | wc -l
For each ~/.claude/skills/*/SKILL.md and ~/.codex/skills/*/SKILL.md,
use the Read tool to inspect the frontmatter (top of file, between --- markers). Extract name and the full description as YAML semantics dictate.
Support all four YAML scalar styles:
| 写法 | 处理 |
|---|---|
单行: description: foo bar | 直接取冒号后内容 |
引号: description: "foo bar" 或 'foo bar' | 去掉首尾引号 |
> folded(多行折叠) | join indented continuation lines with spaces |
| literal(多行保留) | preserve line breaks |
停止条件:遇到下一个未缩进的 frontmatter key(行首无空格且形如 key:),或遇到关闭的 --- 行。如果 description 字段缺失,回落到 <目录名> (no description)。
绝不使用 head \| grep —— 那会把 >/\| 多行风格静默截断到只剩 >,这是 v2.2 之前的真实 bug。务必 Read 完整 frontmatter 后按 YAML 语义解析。
枚举候选 skill 目录:
ls -d ~/.claude/skills/*/ ~/.codex/skills/*/ 2>/dev/null
然后对每个目录:Read 它的 SKILL.md 头部 ~30 行 → 按上表解析 YAML → 输出 <source>|<name>|<full_description>。
记录每个 skill 是「自建」还是「安装」。如果 skill 目录下有 git remote 指向用户自己的 repo,标记为自建;否则标记为安装。
Read ~/.claude/settings.json. Count:
hooks 个数(结构化自动化能力)mcpServers 个数(外部能力接入)permissions.defaultMode~/.codex/)The primary analytics store is ~/.codex/state_5.sqlite, table threads.
Always open with mode=ro so you can never write:
SQ='sqlite3 file:'"$HOME"'/.codex/state_5.sqlite?mode=ro&immutable=1'
# Aggregate
$SQ "SELECT COUNT(*), SUM(tokens_used), MIN(created_at), MAX(created_at) FROM threads;"
# Model breakdown (note: empty/NULL model = older sessions, label as 'Codex (未标注)')
$SQ "SELECT COALESCE(NULLIF(model,''),'Codex(未标注)'), COUNT(*), SUM(tokens_used) \
FROM threads GROUP BY 1 ORDER BY 3 DESC;"
# Reasoning effort distribution (xhigh / high / medium / low / unspecified)
$SQ "SELECT COALESCE(NULLIF(reasoning_effort,''),'unspecified'), COUNT(*) \
FROM threads GROUP BY 1 ORDER BY 2 DESC;"
# Top 15 working dirs
$SQ "SELECT cwd, COUNT(*), SUM(tokens_used) FROM threads \
WHERE cwd != '' GROUP BY cwd ORDER BY 2 DESC LIMIT 15;"
# Hour-of-day heatmap
$SQ "SELECT strftime('%H', datetime(created_at,'unixepoch')), COUNT(*) \
FROM threads GROUP BY 1 ORDER BY 1;"
# Day-of-activity timeseries
$SQ "SELECT date(created_at,'unixepoch'), COUNT(*) FROM threads GROUP BY 1;"
# Sample titles + first user messages for keyword extraction (titles only — no body)
$SQ "SELECT title FROM threads WHERE title != '' ORDER BY created_at DESC LIMIT 200;"
$SQ "SELECT first_user_message FROM threads WHERE first_user_message != '' \
ORDER BY created_at DESC LIMIT 200;"
# CLI versions used (Codex evolution signal)
$SQ "SELECT cli_version, COUNT(*) FROM threads WHERE cli_version != '' \
GROUP BY 1 ORDER BY 2 DESC LIMIT 10;"
# --- 以下为 v2.0 新增查询 ---
# 月度聚合(Evolution 曲线用)
$SQ "SELECT strftime('%Y-%m', datetime(created_at,'unixepoch')), COUNT(*), \
SUM(tokens_used), COALESCE(NULLIF(model,''),'unknown') \
FROM threads GROUP BY 1,4 ORDER BY 1,3 DESC;"
# CLI 版本时间线(Evolution 曲线用)
$SQ "SELECT cli_version, MIN(date(created_at,'unixepoch','localtime')), \
MAX(date(created_at,'unixepoch','localtime')), COUNT(*) \
FROM threads WHERE cli_version != '' GROUP BY 1 ORDER BY 2;"
# 每项目 token 消耗(双工具编排分析用)
$SQ "SELECT cwd, COALESCE(NULLIF(model,''),'unknown'), COUNT(*), SUM(tokens_used) \
FROM threads WHERE cwd != '' GROUP BY 1,2 ORDER BY 1,4 DESC;"
~/.codex/history.jsonl — {session_id, ts, text}. Sample for keywords:
wc -l ~/.codex/history.jsonl # total prompts
jq -r '.text' ~/.codex/history.jsonl | head -300 > /tmp/codex_text.txt # corpus
jq -r '.session_id' ~/.codex/history.jsonl | sort -u | wc -l # distinct sessions
ls ~/.codex/skills/ | wc -l # codex skills
ls ~/.codex/automations/ | wc -l # scheduled automations
ls ~/.codex/rules/ | wc -l # custom rules
~/.gemini/antigravity/)Antigravity is the third local AI tool source. Treat each
~/.gemini/antigravity/brain/<uuid>/ directory as one Antigravity task/session.
Only count directories whose basename is a UUID; ignore non-task directories such
as tempmediaStorage.
Only read local text data:
*.metadata.json for artifact metadata and summariestask.md, implementation_plan.md, walkthrough.md.resolved, .resolved.0, .resolved.1, etc.Never read for analytics:
*.png, *.webp, *.jpg, *.jpeg)~/.gemini/antigravity/annotations/*.pbtxt~/.config/Antigravity/* browser/cache dataAG_BRAIN="$HOME/.gemini/antigravity/brain"
AG_UUID_RE='[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}'
# Count Antigravity task/session directories; exclude temp/media helper dirs
find "$AG_BRAIN" -mindepth 1 -maxdepth 1 -type d 2>/dev/null \
| grep -E "/$AG_UUID_RE$" | wc -l
# Artifact type breakdown from metadata in task/session directories
find "$AG_BRAIN" -mindepth 2 -maxdepth 2 -name '*.metadata.json' -type f 2>/dev/null \
| grep -E "/$AG_UUID_RE/[^/]+\.metadata\.json$" \
| xargs -r jq -r '.artifactType // "unknown"' | sort | uniq -c | sort -rn
# Activity by day from metadata updatedAt
find "$AG_BRAIN" -mindepth 2 -maxdepth 2 -name '*.metadata.json' -type f 2>/dev/null \
| grep -E "/$AG_UUID_RE/[^/]+\.metadata\.json$" \
| xargs -r jq -r '.updatedAt // empty' \
| cut -c1-10 | grep -E '^[0-9]{4}-[0-9]{2}-[0-9]{2}$' \
| sort | uniq -c
# Monthly activity for Evolution curve
find "$AG_BRAIN" -mindepth 2 -maxdepth 2 -name '*.metadata.json' -type f 2>/dev/null \
| grep -E "/$AG_UUID_RE/[^/]+\.metadata\.json$" \
| xargs -r jq -r '.updatedAt // empty' \
| cut -c1-7 | grep -E '^[0-9]{4}-[0-9]{2}$' \
| sort | uniq -c
# Summaries for topic extraction; do not quote full text in the README
find "$AG_BRAIN" -mindepth 2 -maxdepth 2 -name '*.metadata.json' -type f 2>/dev/null \
| grep -E "/$AG_UUID_RE/[^/]+\.metadata\.json$" \
| xargs -r jq -r '.summary // empty' | head -200
# Markdown headings for topic extraction
find "$AG_BRAIN" -mindepth 2 -maxdepth 2 -type f \
\( -name 'task.md' -o -name 'implementation_plan.md' -o -name 'walkthrough.md' \
-o -name 'task.md.resolved*' -o -name 'implementation_plan.md.resolved*' \
-o -name 'walkthrough.md.resolved*' \) 2>/dev/null \
| grep -E "/$AG_UUID_RE/[^/]+$" \
| xargs -r grep -hE '^#{1,3} ' | head -200
# Checkbox volume, useful for task/planning depth
find "$AG_BRAIN" -mindepth 2 -maxdepth 2 -type f \
\( -name 'task.md' -o -name 'implementation_plan.md' -o -name 'walkthrough.md' \
-o -name 'task.md.resolved*' -o -name 'implementation_plan.md.resolved*' \
-o -name 'walkthrough.md.resolved*' \) 2>/dev/null \
| grep -E "/$AG_UUID_RE/[^/]+$" \
| xargs -r grep -hE '^- \[[ xX/-]\]' | wc -l
# Antigravity text artifact scale. This is NOT billing usage and MUST NOT be
# merged into Claude/Codex token totals.
find "$AG_BRAIN" -mindepth 2 -maxdepth 2 -type f \
\( -name 'task.md' -o -name 'implementation_plan.md' -o -name 'walkthrough.md' \
-o -name 'task.md.resolved*' -o -name 'implementation_plan.md.resolved*' \
-o -name 'walkthrough.md.resolved*' \) 2>/dev/null \
| grep -E "/$AG_UUID_RE/[^/]+$" \
| xargs -r wc -l -m \
| awk '
$NF != "total" { files++; lines += $1; chars += $2 }
END {
printf "antigravity_text_files=%d\n", files + 0
printf "antigravity_text_lines=%d\n", lines + 0
printf "antigravity_text_chars=%d\n", chars + 0
printf "antigravity_estimated_token_equivalent=%d\n", int(chars / 4 + 0.5)
}'
Compute:
antigravity_tasks = count of brain/<uuid>/ directories.antigravity_artifacts_by_type = counts by artifactType.antigravity_active_days = unique dates from valid updatedAt values.antigravity_first_active / antigravity_last_active = min/max valid updatedAt dates.antigravity_monthly_activity = monthly counts from valid updatedAt values.antigravity_topics = metadata summaries + markdown headings + checkbox section labels, used only for keywords and high-level themes.antigravity_text_files = count of eligible Antigravity text artifact files.antigravity_text_chars = total character count across eligible Antigravity text artifacts.antigravity_text_lines = total line count across eligible Antigravity text artifacts.antigravity_estimated_token_equivalent = round(antigravity_text_chars / 4) as a rough text-scale proxy only.Antigravity data does not expose verified billing token counts. Use — in token columns or omit token metrics for Antigravity. If reporting antigravity_estimated_token_equivalent, label it exactly as estimated token-equivalent (non-billing) and keep it outside all real token totals, token economics tables, and billing/paid-token claims.
gh)gh auth status >/dev/null 2>&1 || { echo "gh not auth'd, skipping"; }
If authenticated:
# Last-365-day contributions + top repos in window
gh api graphql -f query='
query($from: DateTime!, $to: DateTime!) {
viewer {
login name bio
contributionsCollection(from: $from, to: $to) {
totalCommitContributions
totalPullRequestContributions
totalIssueContributions
totalRepositoryContributions
totalPullRequestReviewContributions
restrictedContributionsCount
contributionCalendar { totalContributions
weeks { contributionDays { date contributionCount } } }
commitContributionsByRepository(maxRepositories: 25) {
contributions { totalCount }
repository { nameWithOwner isPrivate isFork stargazerCount
primaryLanguage { name } }
}
}
repositories(first: 1, ownerAffiliations: OWNER) { totalCount }
pullRequests(first: 1) { totalCount }
issues(first: 1) { totalCount }
}
}' -F from="$(date -u -v -365d +%Y-%m-%dT00:00:00Z)" \
-F to="$(date -u +%Y-%m-%dT00:00:00Z)"
Then page through repositories for language bytes (up to 5 pages × 100 repos):
gh api graphql -f query='
query($cursor: String) {
viewer { repositories(first: 100, after: $cursor, ownerAffiliations: OWNER,
isFork: false, orderBy: {field: UPDATED_AT, direction: DESC}) {
pageInfo { hasNextPage endCursor }
nodes { nameWithOwner isPrivate stargazerCount
languages(first: 10, orderBy: {field: SIZE, direction: DESC}) {
edges { size node { name } } } }
} } }' -F cursor=""
Aggregate languages by Σ size per language across all repos.
Build the candidate path set from:
cwd recovered for each ~/.claude/projects/<encoded>/cwd column from Codex threads tableFor each path that's a git repo, count the current user's commits in the past year:
me=$(git config --global user.email)
for path in <candidate-paths>; do
[ -d "$path/.git" ] || continue
git -C "$path" log --since=1.year.ago --author="$me" \
--numstat --no-renames --pretty=format:'COMMIT|%H|%aI'
done
Aggregate:
commits (count of COMMIT| lines)additions, deletions (sum the numstat columns)last_commit_iso+ - per file extension → top 10 languages)dailyActivity, codex by_date, Claude history by_date, and
antigravity_active_daysmin..max of those datesIf available, include antigravity_text_files, antigravity_text_chars, antigravity_text_lines, and antigravity_estimated_token_equivalent as an Antigravity artifact scale note, not as real token usage.
commits_per_day = git_local_commits / active_daysloc_churn_per_day = (additions + deletions) / active_dayssimultaneous_repos = count of repos with ≥1 commitcross_stack_langs = count of distinct primary languages across repos—,不要估算。antigravity_tasks、artifact type breakdown、
walkthrough / implementation_plan / task artifacts,用来描述「从任务 → 计划 → walkthrough」的交付闭环。名称 | 一句话描述 | 调用次数(如可从 history 统计)。
区分「自建」(用户原创)与「安装」(第三方)。
这一段的叙事重点:不只是 AI 的使用者,更是 AI 工作流的建设者。/plan count / non-command prompt counttotalMessages / totalSessions/plan 开头的 session 占比 → 说明「先想再做」的习惯有多强/compact 或 /clear 的比例 → 上下文管理意识/effort 在 session 内的切换频率 → 是否按阶段调节推理深度/resume 使用率 = resume_count / totalSessions → session 连续性
用 2-3 句话总结出用户的 session 驾驭模式(例如:
「典型流程:/plan 规划 → 迭代 → /compact 回收上下文 → 继续交付」)sessions(claude) + codex_threads + antigravity_tasks +
git_commits + git_linessessions*5 + codex_threads*4 + antigravity_tasks*4 + git_commitsclaude_ratio = claude_sessions / total_ai_unitscodex_ratio = codex_threads / total_ai_unitsantigravity_ratio = antigravity_tasks / total_ai_unitstotal_ai_units = claude_sessions + codex_threads + antigravity_taskscwd basename、GitHub repo 描述 / topics / primary language、Codex thread titles、Claude history first prompts、本地文件名提示(如 package.json 依赖、frontend/、apps/web/、api/)。匿名化只发生在最终输出阶段。deploy、router、ops、docker 等单个工程词,就把一个有明显用户界面或业务功能的产品项目归到“基础设施 / 部署”。| 领域 | 关键词 / 证据(小写匹配 cwd basename + 标题 + 项目信号) |
|---|---|
| 产品 / 业务前端 | frontend, front-end, web, app, h5, mobile, miniapp, ui, ux, page, route, router, dashboard, console, admin, portal, client, website, next, react, vue, vite, svelte, tailwind, shadcn, electron, extension, 小程序, 前端, 页面, 官网, 管理台, 控制台, 后台 |
| 产品 / 业务后端 | backend, server, api, service, gateway, worker, queue, job, db, database, prisma, django, fastapi, express, nest, auth, billing, payment, user, backend service, 后端, 服务端, 接口, 鉴权, 支付, 用户 |
| 产品 / 业务全栈 | product, saas, crm, cms, workspace, studio, platform, marketplace, ecommerce, shop, chat, editor, dashboard + api, web + api, app + server, 产品, 业务, 工作台, 平台, 商城 |
| AI 工具 / Skill | skill, claude, codex, agent, subagent, mcp, prompt, workflow, plugin, antigravity, easy_claude, vibe-forge, readme.skill |
| 基础设施 / 部署 | deploy, infra, ops, monitor, observability, k8s, ci-cd, docker, compose, terraform, nginx, ingress, traefik, caddy, api-gateway, gateway infra, healthcheck, log, cron, 自动化部署, 巡检 |
| 数据 / 分析 | analytics, data, dataset, bi, report, metrics, dashboard analytics, crawler, scrape, readyourusers, bibili, 埋点, 数据, 报表, 分析, 采集 |
| ML / RL / 论文 | rllunwen, rl-, ml, model, training, eval, paper, thesis, 论文, 实验, 大创 |
| 文档 / Markdown | readme, doc, docs, documents, markdown, profile, report, handbook, 文档, 手册 |
| 其他 | 证据不足时的 fallback |
React、Next.js、dashboard、API、billing、agent、deploy。project、repo、test、fix、update、misc、code、task。信号不足,不要补想象中的业务特征。Corpus: 拼接 plan titles + Codex thread titles + first_user_messages
Tokenize:
[A-Za-z][A-Za-z0-9_-]+,小写化,长度 ≥ 2,去停用词[\u4e00-\u9fff]+ 串,做 2-char 滑窗,每个 chunk 内去重;
过滤明显碎片(如"前项""解当""目了"),过滤含纯停用字的二元英文停用词(精简):the, a, an, is, are, of, in, on, for, to, by, from, with, this, that, it, you, we, they, do, does, please, help, use, used, plan, make, get, just, also, will, would, can.
中文停用字:的 了 是 在 和 有 不 就 也 为 以 对 把 被 从 等 都 这 那 个 啊 呢 吧 呀 之 与 或 及 并 要 做 能 会 上 下 里 们 好 之 吗 一 也 就 都 还 到 去 给 跟 向 自 什 么 怎 哪 如 何 因 所 然 后 比 例 而 且 但 不 过 或 还 关 通 基 由 得 着 过 看 想 说 点 种 次 时 年 月 日 中 时 间 现 在.
输出:top 30 keywords,过滤掉只剩 1 出现的,过滤包含纯英文 stop-only 字符的。
用作"标签云"展示:tag1·N tag2·M tag3·K。
hourCounts (claude) + codex by_hour + history by_hour + Antigravity updatedAt hourgh languages.bytes 与本地 git numstat ext 排序commits_per_day = git_local_commits / active_daysloc_churn_per_day = (additions + deletions) / active_daysgithub_contribs_per_active_day = calendar_total / active_daystokens_per_commit / tokens_per_loc)已迁移到 6.10 Token 经济学,6.7 只讲 GitHub / 仓库 / 语言这一维度的目的是回答:如果没有 AI 协作,这种产出可能吗?
计算并叙述:
目的:把静态快照变成成长叙事。让读者看到 AI 使用的成熟度曲线。
数据来源:
threads 的 created_at + model + cli_versionhistory.jsonl 的 timestamp + display(斜杠命令)projects/ 目录的 JSONL 文件创建时间updatedAt + artifactType + summaries计算:
/plan 的日期 → Plan-mode 解锁/effort 的日期 → Reasoning effort 解锁/skill-creator 或自研 skill 出现的日期 → Skill 自建解锁/vibe-forge 或 /ssh-prod 的日期 → 自建 skill 投入生产渲染为 timeline 格式:
2026-01 Codex 起步,纯 prompt,CLI 0.81.0-alpha
2026-02 开始日常化,tokens 增长
2026-03 Claude Code 加入 → plan-mode + effort 调节 → 开始自建 skills
2026-04 双工具编排成熟,skills 生态完善,日均 10 commit
目的:不只展示"用了多少 token",而是讲清 token 投入怎么花、Cache leverage 多深、模型迁移如何省了成本。把 token 当 AI 时代的"原材料 + 杠杆"来叙事,不是产出的注脚。
数据来源:
stats-cache.json 的 modelUsage + dailyModelTokensthreads 的 tokens_used + 月度聚合(Step 3.1 已查)Antigravity estimated token-equivalent (non-billing) is a text-scale proxy from local artifacts. Do not add it to claude_tokens_spent, claude_cache_read, codex_tokens, total token-through, cache leverage, paid/new token totals, or per-model token tables. It may appear only in an Antigravity/local-artifact subsection or a clearly labeled footnote.
计算:
claude_spent + codex_tokens(新付费 token 总量)claude_cache_read(缓存复用 token 总量)claude_cache_read / claude_spent(每 1 个新 token 撬动几个缓存 token)model.spent / Σ all_models.spent(Claude 与 Codex 合在一起算)model.cache_read / model.spent(哪个模型 caching 习惯最熟)dailyModelTokens + Codex 月度聚合按月汇总,找增长拐点<YYYY-MM>: Opus 4.6 spent ↓ 40%,Sonnet 4.6 spent ↑ 60%)claude_spent / git_commits 与 claude_spent / (additions + deletions)叙事重点:
Before writing the README, scan all string fields for these regex and replace
with <REDACTED:type>:
| pattern | replacement |
|---|---|
sk-[A-Za-z0-9_-]{20,} | <REDACTED:openai-key> |
sk-ant-[A-Za-z0-9_-]{20,} | <REDACTED:anthropic-key> |
gh[oprs]_[A-Za-z0-9]{20,} | <REDACTED:github-token> |
github_pat_[A-Za-z0-9_]{20,} | <REDACTED:github-pat> |
AKIA[0-9A-Z]{16} | <REDACTED:aws-key> |
xox[baprs]-[A-Za-z0-9-]{10,} | <REDACTED:slack-token> |
https://open.feishu.cn/open-apis/bot/v2/hook/[A-Za-z0-9-]+ | <REDACTED:feishu-webhook> |
\b[\w.+-]+@[\w-]+\.[\w.-]+\b | <REDACTED:email> |
Project name handling (anonymize=on):
项目 A/B/C/...
in descending order of comprehensive score (Step 6.4)Private Repo Xbasename, never the absolute pathIf user said "show real names" / "私人版":
Choose the profile language before writing:
output/profile_<YYYYMMDD>.md,使用中文叙事output/profile_<YYYYMMDD>_en.md,使用英文叙事Use exactly this structure. For English output, translate headings and prose
to English following examples/profile_20260508_en.md; for Chinese output, use
the structure below. Technical terms stay in English in both versions. 讲故事优先于堆数据;
展示「因为 AI 而不同」,而不仅仅是「用了很多 AI」:
# <name or github_login> · AI-Native Developer Profile
> 基于 <span_days> 天的本地 Claude Code + Codex + Antigravity 数据自动生成 · <generated_at>
> _个人理念:<github bio if available>_
---
## 一览
- 在 **<span_days>** 天里完成 **<claude_sessions>** 次 Claude sessions + **<codex_threads>** 次 Codex threads + **<antigravity_tasks>** 次 Antigravity tasks,共 **<total_messages>** 条 Claude/Codex 消息
- 日均产出:**<commits_per_day>** commits / **<loc_churn_per_day>** 行代码变动 / **<github_contribs_per_day>** GitHub contributions
- 同时维护 **<git_repos>** 个仓库,横跨 **<cross_stack_langs>** 门语言
- 同期 GitHub:**<github_commits>** commits / **<github_prs>** PRs / **<github_issues>** issues / **<calendar_total>** 总贡献
- AI 投入:**<claude_spent>** Claude 新付费 token + **<codex_tokens>** Codex token;复用 **<claude_cache_read>** 缓存(占 Claude I/O **<cache_pct>%**)
- 主力工具:Claude Code (Opus / Sonnet) + Codex CLI (GPT) + Gemini Antigravity
## 🚀 Velocity & Leverage — AI 让一个人拥有了小团队的交付能力
> <1-2 句叙事,例如:「13 个仓库、5 门语言、日均 10 commit —— 这种跨栈广度和交付密度,只有 AI 协作才现实。」>
| 指标 | 数值 | 说明 |
| --- | ---: | --- |
| 日均 commits | <n> | 本地 git commits / 活跃天数 |
| 日均代码变动 | <n> 行 | (additions + deletions) / 活跃天数 |
| 同时维护仓库 | <n> 个 | 过去一年有 commit 的仓库数 |
| 跨栈语言 | <n> 门 | Python / TypeScript / Rust / Go / … |
| GitHub 贡献爆发 | <dates> | 连续 3+ 天 daily > 20 的窗口 |
| 开源影响力 | <total_stars> stars | 跨 <n> 个被 star 的仓库 |
## 🤖 AI-Native 实践
> 不是「偶尔问问 AI」,是把多 LLM 编排、planning、structured workflows 都跑通。
### 多模型编排
| 工具 / 模型 | sessions / threads / tasks | spent tokens | cache-read | 用途倾向 |
| --- | ---: | ---: | ---: | --- |
| Claude Opus / Sonnet | … | … | … | 深度推理、复杂规划、代码修改 |
| GPT (Codex CLI) | … | … | — | 第二意见、跨工具诊断、命令行实现 |
| Gemini Antigravity | <antigravity_tasks> | — | — | 任务制规划、walkthrough、UI/实现闭环 |
| … | | | | |
### 高级能力深度使用
- **Plan-mode**: **<n>** 次
- **Effort 调节**: **<n>** 次
- **Skills**: 共 **<n>** 个(Claude <n> + Codex <m>)
- **Plans**: **<n>** 份;Tasks: **<n>** 个
- **Hooks**: **<n>** 个;Automations: **<n>** 个
- **Antigravity**: **<antigravity_tasks>** tasks;Artifacts: <artifact_type_breakdown>
### Prompt caching 熟练度
每花费 1 个新 token,复用 **<ratio>** 个缓存 token(cache-read 占总 IO 的 **<%>**)。
### Reasoning effort 偏好
xhigh **<n>**(**<%>**)· high **<n>** · medium **<n>** · low **<n>**
## 🔧 AI 基础设施 — 不只用 AI,还在给 AI 造工具
> <1 句叙事:「从 skill 到 hook 到 automation,我在构建让 AI 更好地帮我工作的基础设施。」>
### 自建 Skills
| 名称 | 描述 | 调用次数 | 工具 |
| --- | --- | ---: | --- |
| <skill_name> | <一句话> | <n> | Claude / Codex |
| … | | | |
### 安装的 Skills
| 名称 | 描述 | 工具 |
| --- | --- | --- |
| … | | |
### 其他基础设施
- Hooks: <列出>
- Codex automations: <列出>
- Codex rules: <列出>
## 🛠️ AI 协作风格
### 最常用的 slash 命令 Top 10
| # | 命令 | 次数 | 含义 |
| --- | --- | ---: | --- |
| 1 | /effort | <n> | 切换推理深度 |
| 2 | … | … | … |
### Session 架构
<2-3 句描述用户的 session 驾驭模式,例如:>
- 典型流程:`/plan` 规划 → 深度迭代 → `/compact` 回收上下文 → 继续交付
- **<n>%** 的 session 以 `/plan` 开头(先想再做)
- **<n>%** 的 session 使用过 `/compact` 或 `/clear`(主动管理上下文)
- `/resume` 恢复率: **<n>%**(session 连续性)
- 平均会话深度: **<n>** 条消息 / session
- 最长 session: **<hours>** 小时 / **<msgs>** 条消息
## 📂 项目与领域分布
跨 **<n>** 个项目活跃,按领域分布:
| 领域 | 项目数 | 特征 |
| --- | ---: | --- |
| 产品 / 业务前端 | <n> | React、dashboard、管理台 |
| … | … | … |
### Top 项目(脱敏)
| 项目 | Claude | Codex | Antigravity | Git commits | 编排模式 | 领域 |
| --- | ---: | ---: | ---: | ---: | --- | --- |
| 项目 A | <n> | <n> | <n> | <n> | 三引擎 | … |
| 项目 B | <n> | <n> | <n> | <n> | Antigravity 主导 | … |
| … | | | | | | |
编排模式统计:三引擎 **<n>** 个 · 双引擎 **<n>** 个 · Claude 主导 **<n>** 个 · Codex 主导 **<n>** 个 · Antigravity 主导 **<n>** 个
## 🧬 Evolution 曲线 — AI 用法在进化
<里程碑事件> <里程碑事件> <里程碑事件> <里程碑事件>
月度活跃趋势:
| 月份 | Claude sessions | Codex threads | Antigravity tasks | 里程碑 |
| --- | ---: | ---: | ---: | --- |
| … | | | | |
## 💡 兴趣主题 & 关键词
> **<tag1>** · <tag2> · <tag3> · … (top 25,已过滤停用词)
## ⏱️ 工作节奏
### 24 小时活跃热力图
00 … 01 … …
(峰值时段 / 活跃模式描述)
### 时间跨度
- 首次 / 最近活跃: …
- 活跃天数 / 最长连续 / 单日峰值: …
## 💎 Token 经济学
> 一句叙事开场:「**<spent_tokens>** 新付费 token 撬动 **<cache_read>** 缓存复用,杠杆比 **1 : <leverage>**;总通过我手里 **<total_tokens>** token。」
### 每模型 token 明细(按 spent 排序)
| 模型 | spent | cache-read | leverage | 占总 spent |
| --- | ---: | ---: | ---: | ---: |
| Claude Opus 4.6 | … | … | …× | …% |
| Claude Sonnet 4.6 | … | … | …× | …% |
| Claude Haiku 4.5 | … | … | …× | …% |
| GPT-5.4 (Codex) | … | — | — | …% |
| … | | | | |
### 月度 token 趋势
| 月份 | Claude spent | Claude cache | Codex tokens | 主力模型 | 注解 |
| --- | ---: | ---: | ---: | --- | --- |
| <YYYY-MM> | <n> | <n> | <n> | <model> | <事件 / 迁移> |
### 模型迁移注解
- <YYYY-MM>: <模型 A 萎缩 −X%>,<模型 B 接管 +Y%>,<推测原因>
- …
### 单位投入产出(仅参考,勿当 KPI)
- 每 commit ≈ **<n>** Claude tokens(仅 spent,不含 cache 与 Codex)
- 每行代码 ≈ **<n>** Claude tokens
> 提醒:AI 产出还包含大量不直接转化为 commit 的高价值劳动(架构 review / 数据清洗 / plan 推演 / skill 重构)。把"每 commit X tokens"当 KPI 是反激励。
## 💰 产出 & 投入
### GitHub 同期产出
- 365 天总贡献: **<n>** · 拥有仓库: **<n>**
- 最高产单日: <top 5 dates>
#### Top 仓库
| 仓库 | language | commits | stars |
| --- | --- | ---: | ---: |
| … | | | |
### 主要语言
<语言列表>
## 📊 数据来源 & 隐私承诺
- 数据 100% 本地:`~/.claude/*` + 项目 `.claude/plans`(如配置) + `~/.codex/*` + `~/.gemini/antigravity/brain/*` + 本地 `git log` + GitHub via `gh`
- Claude plans 同时覆盖默认 `~/.claude/plans` 与 settings 中解析出的 `plansDirectory`
- 对话正文仅用于关键词与协作风格分析,Antigravity 只读取 metadata summary 与 markdown headings/checkbox,不读取截图、浏览器 cache 或 pbtxt annotations;原文不会出现在报告中
- 项目名已匿名,API key / token / 邮箱 已正则清洗
- 报告由 Claude Code / Codex / Antigravity 本地数据按 Readme.skill 自动生成,可重复运行
- 生成时间: **<ISO timestamp>**
如果用户说"生成海报" / "AI 海报" / "social card" / "可分享图" / "make poster",或默认就把它当一份附加交付物,在 markdown profile 完成后再渲染 SVG 海报到 output/poster_<YYYYMMDD>_<lang>.svg(例如 _zh.svg / _en.svg)。
font-family="system-ui, -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif",不内嵌字体—,不补想象指标| 区域 | y 位置 | 内容 |
|---|---|---|
| 顶部品牌条 | 120 | 6px 渐变小条(accent) |
| 一句叙事标题 | 200–345 | 标签 + 大字标题 + 时间跨度副标题 |
| 6 hero metric 卡 (2×3) | 440–1140 | 每张卡 465×220 圆角,数字 100-120px,标签 20px letter-spaced |
| Evolution timeline | 1240–1430 | 横向 4 milestone(圆点 + 月份 + 事件) |
| 副信息卡(左右两栏) | 1500–1720 | Cache leverage 排行 / Top slash 命令 |
| 底部 footer | 1790–1825 | 脱敏标识 + repo URL + 日期 |
| 卡 | 数据来源(从 10 维度取) |
|---|---|
| 1 | <span_days> · <active_days> ACTIVE(一览) |
| 2 | <git_total_commits> LOCAL COMMITS(Velocity) |
| 3 | <total_through> TOKENS THROUGH(Token 经济学,spent + cache_read 总量) |
| 4 | 1 : <cache_leverage> CACHE LEVERAGE(Token 经济学) |
| 5 | <total_stars> GITHUB STARS(投入产出) |
| 6 | <n_repos> · <n_langs> REPOS · LANGS(Velocity) |
任一项缺数据时,替换为:总 AI units(<claude_sessions> + <codex_threads> + <antigravity_tasks>)/ 自建 skills 数 / Antigravity artifacts 数 / 单日峰值消息数。
#0c0a1f → 中紫 #1a1442 → 深绿 #0d2e1f#8b5cf6 → Codex 绿 #10b981<linearGradient> 定义在 <defs>,fill="url(#bg)" 引用<rect rx="20"> 圆角;分隔线用 <line stroke-opacity="0.1">海报有中英两个版本。决定哪种:
output/poster_<DATE>_zh.svgoutput/poster_<DATE>_en.svg保留英文不翻译(中英版都用英文,因为这是行业标准术语 / 设计感词):
token / tokens / through / cache / leverage / commits / stars / repos / langs / days / active / models / sessions / threadsOpus / Sonnet / Haiku / GPT-5.4 / GPT-5.5 等/effort / /usage / /plan / /compact 等GitHub / Readme.skill / 版本号LOCAL COMMITS / TOKENS THROUGH / EVOLUTION) —— 设计语言,两版都用英文翻译的部分(中英文版差异点):
| 元素 | 中文版 | 英文版 |
|---|---|---|
| 主标题(一句叙事) | 「118 天 · 双引擎 · 一个人的小团队」 | "118 Days · Two-Engine · One-Person Team" |
| Evolution 节点描述 | 「Codex 起步」「tokens ↑6×」「Claude 加入」「双引擎峰值」 | "Codex starts" / "Tokens up 6×" / "Claude joins" / "Two-engine peak" |
| 副信息卡 section 名 | 「CACHE LEVERAGE 排行」「TOP SLASH 命令」 | "CACHE LEVERAGE RANK" / "TOP SLASH COMMANDS" |
| Footer | 都用英文(设计感) | 都用英文(设计感) |
为了让海报有"想转发、想晒、看到的人想自己也来一份"的传播力,海报必须包含以下 3 件套:
不是堆数据,而是让 AI 看了用户数据后,写一段有破圈传播力的评语作为海报副标题。默认用 Tone A(反差数字 + 通俗类比)—— 把 token 量换算成"等于 N 遍世界名著",让圈外人 3 秒被震撼。
把 total_through(spent + cache_read)换算成大众能感知的「读了 N 遍《红楼梦》/ N 倍 War & Peace」。
换算公式:
chinese_chars ≈ total_through × 0.7(1 token ≈ 0.7 个汉字)dhm_count ≈ chinese_chars / 730_000(《红楼梦》约 73 万字)english_words ≈ total_through × 0.75(1 token ≈ 0.75 个英文 word)wap_count ≈ english_words / 587_000(War & Peace 约 58.7 万 words)样例(基于 12.9B token through 的 demo):
117 天,我和 AI 写下 120 亿字 / 等于把《红楼梦》写了 1 万遍117 days · 12.9B tokens with AI / That's War & Peace × 25,000 times为什么 Tone A 优先:12.9B 这种数字对圈外人是抽象的;红楼梦/War & Peace 任何受过教育的人都立刻有量感。这是从「圈内炫耀」变「破圈震撼」的关键。
| 画像(命中即触发) | tone | 中文样例 | 英文样例 |
|---|---|---|---|
| 最长 session messages > 1000 | B 拟人化关系 | 跟 AI 吵了 轮 / 没分手 | -message marathon / Still together |
| commits / LOC 极高 + 跨多 repo | C 角色反转 | 我不再写代码 / 我让代码自己长出来 | I don't write code / I grow code from prompts |
| Cache leverage > 25× | D 自嘲 humble brag | 不是我手快 / 是 Claude 24h 陪我 | I'm not fast / Claude never sleeps |
| token + commits 都极高("打工人") | E 反差悖论 | 老板以为我在摸鱼 / 我和 AI 烧了 B token | Boss thinks I slack / Burned B tokens |
| plan-first 高 + 多自建 skills | F 哲学/思考 | 我不写代码 / 我编排 AI 替我写 | I don't write code / I orchestrate AI |
规则:
total_through < 1B(数字撑不起类比),否则永远先用 Tone Afill="url(#accent)" 渐变色填充,制造视觉重音("红楼梦"那行用渐变)<X> tokens · <Y>× cache · <Z> skills · <N> langs基于数据自动判定徽章。每个用户最多展示 4 个最强徽章(按下表优先级取前 4):
| 徽章 | 触发条件 | 显示文本 |
|---|---|---|
| TWO-ENGINE PILOT | Claude sessions ≥ 50 且 Codex threads ≥ 50 | TWO-ENGINE PILOT |
| THREE-ENGINE PILOT | Claude sessions ≥ 50 且 Codex threads ≥ 50 且 Antigravity tasks ≥ 20 | THREE-ENGINE PILOT |
| CACHE MASTER | claude_cache_leverage ≥ 15× | CACHE MASTER · <leverage>× |
| SKILL BUILDER | 自建 skills + automations + rules ≥ 5 | SKILL BUILDER · <n> |
| POLYGLOT | 跨栈语言 ≥ 5 | POLYGLOT · <n> |
| VELOCITY KING | 日均 commits ≥ 8 | VELOCITY KING · <n>/d |
| PLAN-FIRST | session-first 是 /plan 的占比 ≥ 8% | PLAN-FIRST · <%> |
| WALKTHROUGH BUILDER | Antigravity walkthrough artifacts ≥ 10 | WALKTHROUGH BUILDER · <n> |
| TOKEN WHALE | total_through ≥ 10B | TOKEN WHALE · <total> |
| OPEN-SOURCE | total stars ≥ 1000 | OPEN-SOURCE · <stars>★ |
| EARLY ADOPTER | 使用过 ≥ 3 个不同模型版本 | EARLY ADOPTER |
| LONG-CONTEXT PRO | 用过 Opus 4.7-1M ≥ 10 次 | LONG-CONTEXT PRO |
视觉:圆角胶囊 235×60 (rx=30),1.5px accent 渐变描边,文字 17px letter-spaced 1.5。 4 个胶囊一行排列,gap 25px,左 60px 起。徽章文字两版一致(都用英文,都是设计语言)。
替代单纯 footer,给一个行动召唤区。看到海报的人能直接看到 install 命令。设计:
GENERATE YOURS IN 30 SECONDS(两版都英文,保持设计感)fill="url(#accent)" 突出):
/plugin marketplace add study8677/Readme.skill/plugin install readme-skill@study8677github.com/study8677/Readme.skill(letter-spaced 2px,20px)LOCAL-ONLY · ANONYMIZED · v<version> · <date>(13px,30% opacity)为什么不放二维码:QR 在小屏幕扫描成功率低;让人看到命令直接复制粘贴更可控;保持视觉简洁。让"想生成自己的"的人主动去 google 搜 repo,反而过滤出真正动机强的种子用户。
examples/example_poster_zh.svgexamples/example_poster_en.svg两份对照来看一下"哪些翻译、哪些保留"的具体边界,以及徽章 + 金句 + CTA 在 SVG 里的实现方式。
把 Markdown 写入对应语言的 output/profile_<YYYYMMDD>.md 或 output/profile_<YYYYMMDD>_en.md,
然后给用户一句话总结:
✅ Profile generated: output/profile_<YYYYMMDD>.md (或 output/profile_<YYYYMMDD>_en.md)
🎨 Poster: output/poster_<YYYYMMDD>_<lang>.svg (如果生成了)
关键数字:<claude_sessions> Claude sessions / <codex_threads> Codex threads / <antigravity_tasks> Antigravity tasks / <tokens> tokens / <github_commits> commits
预览:head -40 <profile_path>
预览海报:open output/poster_<YYYYMMDD>_<lang>.svg
转 PNG:rsvg-convert -h 1920 output/poster_<YYYYMMDD>_<lang>.svg > poster.png
(或 chromium --headless --screenshot=poster.png poster.svg)
如果用户要求"私人版",再生成一份 output/profile_<YYYYMMDD>_private.md
跳过项目匿名(仍然 scrub 密钥与邮箱)。
| 缺失项 | 你应该做什么 |
|---|---|
~/.claude/stats-cache.json 不存在 | 跳过 Claude 总量章节,仅基于 history.jsonl 估算 |
~/.codex/state_5.sqlite 不存在 | 跳过 Codex 章节;如果 history.jsonl 仍在,至少给个总数 |
~/.gemini/antigravity/brain 不存在 | 跳过 Antigravity 章节;总览和项目表不显示 Antigravity 列或显示 0 |
| Antigravity metadata 缺失 | 用 markdown 文件名、标题和文件 mtime 降级;count-based metrics 保留,date-based metrics 跳过无效记录 |
| Antigravity 只有截图/二进制 | 只计 brain 目录为 task/session,不读取图片,不做 OCR,不编 topic |
| Antigravity token 不可得 | token 表显示 — 或省略 Antigravity token;不要估算 |
gh 未安装 / 未认证 | 跳过 GitHub 章节,profile 仍可生成 |
| 候选路径不是 git 仓库 | 该项目从 git 统计中跳过 |
| 数据全空 | 报告诚实地说明"暂无可统计的本地数据",不要编数据 |
~/.claude/projects/*/<id>.jsonl 里的 message.content 用于关键词提取、协作风格、Session 架构等深度分析(Step 6.3 / 6.5 受益);但不要把任何对话原文一字不差地写进 README——脱敏后的统计、概括、片段化关键词可以gh 调用 GitHub 自身)~/.claude 或 ~/.codex 下任何文件~/.gemini/antigravity/brain/* 下的 metadata 与 markdown 文本;不要读取 screenshots、pbtxt annotations、~/.config/Antigravity cache,且不要 OCR 图片~/.claude/skills/deploy/SKILL.md — 编号步骤 + bash 示例的简洁风格~/.claude/skills/ops-report/ — 只读 sqlite 查询的范式~/.claude/skills/log-patrol/ — 跨数据源汇总并出表格的范式npx claudepluginhub study8677/readme.skill --plugin readme-skillGenerates brag documents from GitHub activity using LLM APIs (OpenAI/Anthropic). Useful for performance reviews, promotion packets, and self-assessments.
Analyzes Claude Code usage patterns from history, detects project tech stacks, discovers GitHub community resources, suggests CLAUDE.md improvements, and generates agent/skill/command configs.
Synthesizes persistent user identity profiles from behavioral observations across sessions. Useful for tracking communication styles, decision patterns, and self-reflection with evidence-backed claims.