God Skills

⚡ Run God Skills in Zokie, the IDE built for them. Download Zokie →

Seven Claude Code skills that behave like one company instead of one assistant guessing outside its expertise — a CEO, a builder, a QA gate, a CFO, a PM, a writer and a wellbeing assistant. Each owns a domain, hands work to the next, and learns from every run: a lesson that would help every user becomes a pull request to this repo; a lesson about you stays on your machine.

xhanthis.github.io/god-skills — why the seven exist, how a request flows through them, and a router you can try.

Two packages ship from this repo:

| Package | What it installs | |---|---| | god-skills | the 7 skills — knowledge, loaded into your session | | god-agents | subagents generated from those skills, hook gates that enforce them, the god-ally activity hook, and an unattended runner (docs) |

Install

npx god-skills            # every skill, asks global or project
npx god-skills list       # see the 7
npx god-skills doctor     # verify the install, flag leftovers from older versions

npx god-skills dev qa --global      # just these two
npx god-skills --project --force    # overwrite in this repo
npx god-agents --all                # subagents, /god, and the hook gates

| Flag | Does | |---|---| | -g, --global | installs to ~/.claude (every project) | | -p, --project | installs to ./.claude (this repo, checked in for the team) | | -a, --all | every skill, no prompt | | -f, --force | overwrite files already there | | -y, --yes | no prompts, defaults to global |

Short names work: npx god-skills dev installs god-dev. Upgrading from 2.x: the install removes the retired skill folders it once wrote (never a folder of yours that happens to share a name) and doctor names any left behind. v2.4.0 is tagged if you need the old 30.

Restart Claude Code after installing — skills load at session start.

Other CLIs — Codex, Gemini, Cursor, Copilot

The skills are plain Markdown, so any agent that reads an instruction file can follow them:

npx god-skills --codex                            # ./AGENTS.md   (OpenAI Codex CLI)
npx god-skills --gemini                           # ./GEMINI.md   (Gemini CLI)
npx god-skills --agents-md .cursor/rules/god.md   # any instruction file

Each copies the 7 skills to ./.god-skills/ and adds one marker-delimited index block to the file (re-runs replace the block; the rest of the file is untouched). What does not port: the hook gates (Result: PASS before a session can end), the auto-trigger card, and the Claude Docs / gstack browse integrations — those are Claude Code features. The agent reads a skill when its trigger matches and runs them in sequence itself.

The seven

| Skill | Core question | Absorbed | |---|---|---| | god-ceo | What is the real problem, is it worth doing, who does it, and what is the final call? | context, cos, historian, da, god | | god-dev | How do we design and implement this correctly, fast, and better than last time? | architect, simplifier | | god-qa | Does it actually work, is it safe, and can we prove it? | tester, security, police, pl | | god-cfo | Do the numbers reconcile, what should we charge, and what does the data say? | pricer, data | | god-pm | What should we build, why, for whom, how does it run, and how is the competitor built? | customer, ops, strategist, scout, researcher, reverse | | god-cmo | Is this clear, short, and does it read like a human wrote it? | write, editor | | god-ally | Is this pace sustainable — and should you be working right now? | plan, health |

Each replies in its own shape, in plain Markdown that reads the same in a terminal, in chat and in an IDE: god-dev a build report (title, one-line summary, the links, then QA and what needs you), god-qa a verdict with a checklist and an issues table, god-ceo a decision memo with a before/after, god-cfo a worked example in tables, god-pm a customer moment with sourced evidence, god-cmo the rewrite itself with a short edit note. god-ally alone prints a fenced report, so its bars keep their columns.

Every skill's SKILL.md stays under 150 lines; the heavier passes live in references/ and load only when the task needs them (god-qa's four-viewport frontend pass only when UI changed, its security pass only when auth or input is touched, god-dev's architecture pass only in deep mode).

God Ally's daily report

/god-ally prints one screen: a Zen Score out of 10 for today, a 30-day graph, what you spent, and one thing to change tomorrow. It never asks you a question — a source that isn't there is listed as missing and the score is computed without it.

node ~/.claude/skills/god-ally/scripts/zen-report.js        # or just ask for /god-ally

| Component | Weight | Full marks | |---|---|---| | Day length | 35% | first to last activity within 8 hours | | Sleep | 30% | 7h30m asleep and nothing logged after 21:00 | | Intensity | 20% | tokens and commits at or under your own 28-day median | | Recovery | 15% | two days off in the last seven |

Any activity past midnight, or a night under five hours, caps the score at 5. 8–10 Balanced · 5–7.9 Stretched · under 5 Burnout risk. The day runs 05:00 → 05:00 in Asia/Kolkata, so a 00:29 commit counts against the day before — change both in the config.

Where it reads from, timestamps and counts only, never content: your git commits across the repo roots in the config; Claude Code token usage via npx ccusage@latest daily --json, falling back to ~/.claude/projects/**/*.jsonl; Apple Health sleep exported to iCloud; and, when they are connected, Google Calendar, Slack, Gmail, Linear and Notion through MCP — the skill fetches those timestamps and hands them to the script. Config and history live in ~/.god-ally/, created on the first run, which also backfills 30 days.

Keep enough history. Claude Code deletes its logs after 30 days. Raise it in ~/.claude/settings.json so the baseline has something to stand on:

{ "cleanupPeriodDays": 120 }

Sleep, from your watch. The script reads ~/Library/Mobile Documents/com~apple~CloudDocs/GodAlly/sleep/YYYY-MM-DD.json, each holding {"date":"2026-09-22","asleep_minutes":412,"bedtime":"01:10","wake":"08:02"}. Build it once with an iOS Shortcut, on the iPhone that has your Health data:

1. Shortcuts → Automation → + → Time of Day. Pick a time after you normally wake (09:00 works), Daily, and turn on Run Immediately so it never asks.
2. Add Find Health Samples → type Sleep Analysis, Sort by Start Date, and set the date range to today.
3. Add Calculate Statistics (or Get Numbers from Input) to total the asleep minutes, then Text to build the JSON above — the sample's start is your bedtime, its end your wake time.
4. Add Save File, destination iCloud Drive → GodAlly → sleep, filename <today's date, yyyy-MM-dd>.json, Overwrite if file exists on.

Skip it and the report still works: sleep is then scored on when you stopped working alone, and the footer lists apple health sleep as missing.

The graph is the last seven days, one bar per day: the y-axis is the score out of ten, so taller is a better day, and the date and that day's score sit under every bar. A day off has no bar. It is drawn by the script itself — nothing to install.

A rare word of encouragement. The report itself is numbers and one action, never a quote. But at the close of another skill's reply, --quote will occasionally return a single line — only when the day scored under 6, none has been shown in three days, and a one-in-twenty draw lands. Most calls print nothing, which is the point. There is no quote list: the script describes the moment, the skill writes a fresh line for it, and every line shown is remembered so none repeats.

node ~/.claude/skills/god-ally/scripts/zen-report.js --quote

Keep enough history. Claude Code deletes its logs after 30 days. Raise it in ~/.claude/settings.json so the baseline has something to stand on:

{ "cleanupPeriodDays": 120 }

Sleep, from your watch. The script reads ~/Library/Mobile Documents/com~apple~CloudDocs/GodAlly/sleep/YYYY-MM-DD.json, each holding {"date":"2026-09-22","asleep_minutes":412,"bedtime":"01:10","wake":"08:02"}. Build it once with an iOS Shortcut, on the iPhone that has your Health data:

1. Shortcuts → Automation → + → Time of Day. Pick a time after you normally wake (09:00 works), Daily, and turn on Run Immediately so it never asks.
2. Add Find Health Samples → type Sleep Analysis, Sort by Start Date, and set the date range to today.
3. Add Calculate Statistics (or Get Numbers from Input) to total the asleep minutes, then Text to build the JSON above — the sample's start is your bedtime, its end your wake time.
4. Add Save File, destination iCloud Drive → GodAlly → sleep, filename <today's date, yyyy-MM-dd>.json, Overwrite if file exists on.

Skip it and the report still works: sleep is then scored on when you stopped working alone, and the footer lists apple health sleep as missing.

How a request flows

USER → god-ceo (vague or multi-skill asks only) → god-dev → god-qa → EXECUTE

• Vague request ("order amount is wrong") → god-ceo investigates the codebase, reconstructs the real problem, says whether it is worth doing and where it ranks this week, and picks the minimum chain. A clear single-skill ask skips it.
• Writing code → god-dev picks a mode — small (≤30 lines, no revenue / business logic / data), normal, deep (anything touching money, data, auth, or >300 lines; designs first and waits for your OK) — you can override with one word. It syncs git, checks for an unmerged branch already doing the work, loads the repo profile and lessons, plans with a what-could-go-wrong list, removes before it adds, self-scores on god-qa's 1–5 scale and fixes every 4–5, then runs god-qa itself in the same mode — every change is tested by default — and returns one message: title, a plain-English sentence, PR links, the flowchart doc, the test cases, the manual checks.
• Proving it → god-qa extends the same risk list, writes and runs tests (backend; frontend at 390×844, 820×1180, 1512×982, 1440×900 via gstack browse; accessibility; perf and a 50-call burst on local/staging only; simulated low network), runs the security and Indian-compliance passes when the diff touches them, re-runs past failures first, spot-checks its own evidence, scores every issue (5 = cannot go live … 1 = backlog; any 5 or two 4s = FAIL), fixes and retests up to three times, and ends with a one-screen verdict plus two Claude Docs — the test cases with results, and a manual guide with ready-to-run curls.
• Money in the diff → god-cfo recomputes it a second way before it is trusted.
• Any significant decision → god-ceo attacks it, decides (BUILD · MODIFY · CHEAPEST TEST FIRST · DEFER · INVESTIGATE · DO NOT BUILD · SHIP · DO NOT SHIP · STOP), records why in docs/decisions/, and answers "why does this exist" later.
• Every skill's final message ends with god-ally's status callout — Zen Score against the week, intensity against the month, dollars burned today and over 30 days (zen-report.js --line) — and one more 🧘 line whenever it has something specific: stop for tonight, take a break, don't start this before the meeting, eat, go deep now. On a strong signal it asks before you continue.

Failure loops: qa → dev → qa (max 3) · cfo → dev → cfo · stuck anywhere → god-ceo (sensei).

The learning loop

One shared loop, god-ceo/references/learning-loop.md, used by all seven. A learning is captured from four triggers — you correcting the skill, another skill failing its work, an outcome later (a reverted PR, a wrong decision), and a one-line self-review. It is then judged on who it is true for, not on how often it repeats:

| Scope | Test | Goes to | |---|---|---| | universal | would hold for a stranger in another repo at another company | lessons/global.md; may become a PR here | | repo | true only in that codebase | lessons/<repo>.md, local | | personal | your taste, tools, style | lessons/personal.md, local — even after 100 sightings |

Value = severity × reach × confidence. A universal lesson scoring ≥ 18 (or any severity-5 finding) opens a PR on its first sighting; 6–17 is stored and promoted as confidence grows; below 6 waits as a candidate. The PR carries a summarized rule only — never code, paths, names or tokens — is opened from the user's fork, labelled learning, at most one a day, and is never merged by a machine. On by default; "share_learnings": false in ~/.claude/god/config.json turns it off. CI rejects a learning PR that touches anything but skill text, exceeds 40 lines, lacks its Scope: / Score: lines, or contains anything identifying.

Memory lives in ~/.claude/god/<skill>/ — lessons, scorecards, god-dev's repo profiles, god-qa's regression files, god-ceo's decisions and priorities, god-ally's activity log. Nothing there is ever sent anywhere except the summarized rules described above. The only documents any skill publishes are god-dev's flowchart + change log and god-qa's test cases and manual guide.

The rules every God obeys

No fabrication. Numbers come from the repo, your analytics, or a named public source. Unknown impact is reported as unknown, plus what to instrument to find out. Legal sections and citations are verified, never remembered.

Proof over claims. Tests that could not run return UNVERIFIED, never PASS. A screenshot that was not looked at is not a test.

Adversarial before final. Nothing significant reaches a decision without god-ceo trying to destroy it first, and nothing reaches a PASS without god-qa's what-could-go-wrong list.

Integrity gate. god-qa's integrity pass re-runs one test, re-verifies one claim, and greps the final diff for what the org PR reviewer hard-blocks before any PASS — sampling catches most shortcuts.

Ship gate. god-dev writes to, and god-qa fails on, the org PR auto-reviewer's own checklist, line by line: raw errors in responses, unindexed or unbounded SQL, PII in logs / URLs / third parties / error bodies, missing authz, XSS, secrets and credential-shaped fixtures, the injection markers, size caps, red CI. The gate is met, never gamed — a rule that cannot be met is written into the PR body instead.

Brevity. Lead with the finding, one line per point, max three points outside the templates, no prose on a pass. If the review is longer than the change, the review is wrong.

Hooks (god-agents)

| Event | Gate | |---|---| | Stop | the session cannot finish while god-dev edits lack a god-qa Result: PASS — read from the chain log for subagents and the transcript for inline skills | | SubagentStop on god-qa | the verdict is recorded so the Stop gate has ground truth | | PreToolUse on Edit/Write | string-concatenated SQL and reviewer tripwires are blocked before they land | | SessionStart / UserPromptSubmit / Stop | god-ally's collector logs the event and, before a prompt, warns when a meeting is minutes away or you are past your stop time |

Proof

npm test runs 320 assertions across both packages — no credentials, no network — and CI runs them on every pull request:

| Suite | Covers | |---|---| | skills.test.sh | installer, force and retire semantics, doctor, tarball, the --codex target, every skill's contract (verdict tokens, viewports, modes, memory paths, ≤150-line cores) | | agents.test.sh | generation, tool boundaries, settings-merge safety, retired-agent cleanup, doctor | | hooks.test.sh | every gate incl. the god-ally collector, both jq and python3 paths, fail-open behaviour | | linear.test.sh | the dedup protocol against a mock Linear server | | runner.test.sh | runner guardrails against real throwaway git repos | | zen.test.sh | god-ally's scoring units, the 3.3 worked example, the day boundary, the caps, the fresh-line brief, and a full run with every source missing |

PLAN.md is the original 30-skill design and is kept for history.

License

MIT. god-cmo's pattern catalog is adapted from blader/humanizer (MIT).