NoobyGains/godmode: Engineering discipline for AI coding agents: question the request, verify against current docs, plan, build test-first on the cheapest capable model, review with a second model family, prove it before saying done. One repo for Cla

Your coding agent questions the request, verifies against current docs instead of its memory, plans, builds test-first on the cheapest capable model, gets a second model family to review, and proves the result before it says done.

Install • Demo • Why • How it works • Skills • Harnesses • Compared with • Changelog

What a session looks like

_{An illustration of a session, not a transcript. Every line maps to a skill below.}

Install

Claude Code

/plugin install godmode --marketplace NoobyGains/godmode

On older Claude Code versions use two steps: /plugin marketplace add NoobyGains/godmode, then /plugin install godmode@godmode. Start a new session afterwards: the bootstrap is injected at session start. /godmode:doctor confirms the install.

Antigravity

agy plugin install https://github.com/NoobyGains/godmode

Codex

codex plugin marketplace add NoobyGains/godmode
codex plugin add godmode@godmode

Any agent that reads skills (Cursor, Gemini CLI, OpenCode, Copilot CLI, Windsurf and others)

npx skills add NoobyGains/godmode

This installs the skills into the agent's skills directory. The bootstrap that makes them trigger automatically needs the harness-specific route in the harness table.

[!TIP]

Nothing to configure. GodMode is Markdown skills plus two small hooks (a session-start bootstrap and a one-line reminder per prompt); it never calls a network service, never collects telemetry, and has zero runtime dependencies.

Why GodMode

Without and with

How it works

GodMode is a gated pipeline. The agent cannot move to the next phase until the current gate has evidence.

1. Define. rationale challenges the request. intent-discovery reads the project before its first question, asks one question at a time, presents two or three approaches, writes the design document and has a fresh builder check it. grounding verifies every fact that can change.
2. Plan. reference-first finds the proven implementation to build on. task-planning writes tasks of two to five minutes with exact files, signatures, the tests as code, commands and expected output, plus the interfaces each task consumes and produces.
3. Execute. delegated-execution runs the plan with a fresh subagent per task, a ledger on disk that survives context loss, and a bounded fix loop; inline-execution is the cheap single-context mode; parallel-execution and team-orchestration cover independent and collaborating work. Every task is test-first.
4. Review. code-review runs spec compliance then quality on each task; cross-model-review brings in another model family before landing; review-response handles feedback with evidence instead of reflexive agreement.
5. Ship. quality-gates are machine-checked, completion-gate demands fresh output (on Claude Code it also arms a one-shot stop check that reads the final message and sends an unsupported claim back), merge-protocol lands the branch and cleans up the worktree.

Grounding: the skill nobody else makes the headline

Execution routing and model routing

Agent teams are the most expensive mode, so they are a decision, not a default: the lead enumerates the tasks, maps which ones need information from each other, and only forms a team when teammates must coordinate mid-work. On harnesses without teams the same skill degrades to parallel subagents.

The 33 skills

Every skill has a Prime Directive (the rule that cannot be broken), a Protocol, Cognitive Traps (the rationalisations it blocks) and Guardrails. Bodies describe actions, never one harness's tools; skills/using-godmode/references/ maps those actions to each harness.

Harnesses

| Harness | Install | Bootstrap at session start | Status | |---|---|---|---| | Claude Code | one-step /plugin install line above | SessionStart hook, plus a one-line reminder on each prompt | Verified on 2.1.310: hook tested on sh, with an empty PATH and through cmd.exe; in a clean profile (only Claude Code's bundled plugins present) a headless "Let's make a react todo list" invoked rationale first and answered with options, no code; completion-gate's stop check registers, evaluates and removes itself (evidence). In the default permission mode the first completion-gate call asks to execute the skill because it arms that check; answer once | | Antigravity CLI (agy) | agy plugin install | Always-on plugin rule plus the skill index | Verified on agy 1.4.7 (Windows): installed as a native plugin with all 33 skills; a headless "Let's make a react todo list" on Gemini Flash answered with the rationale analysis, no code (evidence) | | Codex CLI and App | codex plugin marketplace add + codex plugin add | Native skill discovery ($godmode:using-godmode); doctor is the onboarding skill | Verified on Codex CLI 0.163: marketplace registered from this repo and godmode@godmode installed and enabled, with the manifest in the extensions.com.openai shape the Codex plugin docs specify (evidence) | | Gemini CLI | gemini extensions install | Extension context file | Manifest shipped, untested | | Cursor | plugin from repository | sessionStart hook | Manifest shipped, untested | | OpenCode V1 and V2 | opencode.json plugin entry | V1: injected into the first user message; V2: session context hook | Plugin shipped and module-loads under Node; untested on a live install | | Copilot CLI, Windsurf, others | npx skills add NoobyGains/godmode | None (skills trigger from their descriptions) | Skills only |

"Untested" means no maintainer has run it on a live install yet. If you do, open an issue with the transcript and the row changes. The per-prompt reminder is a UserPromptSubmit hook, which fires in interactive sessions; in our headless claude -p runs it did not fire, so the session-start bootstrap alone produced the verified behaviour.

Limits worth knowing: agent teams exist only on Claude Code and are experimental and opt-in there; Antigravity has no todo tool, so plans are tracked in a task artifact; OpenCode V2 has no todo tool either; Codex needs multi-agent enabled for subagent dispatch. Each references/-tools.md spells out the substitutions.

Compared with

Plenty of skill packs teach an agent to design before it codes and to review before it ships. GodMode is built around proof. Every fact that can change is checked against a current source before the agent acts on it; every dispatch names the cheapest model that can do the job; nothing is called done without fresh output and a second model family's review. Where a neighbour is the better pick, the table says so.

| Project | What it is | Use it instead when | |---|---|---| | superpowers | 15 tightly tuned skills, 17 harnesses, an eval lab and a Discord | You want the smallest core with the longest track record | | compound-engineering | Thirty-six skills with a learning loop that writes solutions back into the repo | You want every fix to compound into a project knowledge base | | spec-kit | A CLI-driven spec-first workflow with verdict states | Your team wants formal specifications as the primary artefact | | agent-skills | 25 lifecycle skills installed with npx skills add | You want a lighter set without a bootstrap hook | | GodMode | 33 skills, 6 harness manifests, web-grounded facts, per-seat model routing, cross-model review | You want the agent to verify instead of remember, spend tokens deliberately, and prove results |

Evals

The repository ships a claude plugin eval suite (evals/): five prompts a disciplined agent should handle differently, each scored by regex, file and LLM graders, and run with the plugin and without it, so the delta is the plugin's effect. Numbers below are the mean of three runs per arm.

| Case | With GodMode | Without | Delta | Note | |---|---|---|---|---| | design-before-code ("Let's make a react todo list") | 1.00 | 0.67 | +0.33 | Options and questions every run; the baseline wrote scaffolding in two of three | | grounding-before-sdk-code | 0.67 | 0.17 | +0.50 | Current docs were checked in every run with the plugin, in one of three without; the final message still often omits the citation | | root-cause-before-fix | 1.00 | 0.83 | +0.17 | The baseline is already careful here | | no-done-without-evidence | 1.00 | 1.00 | 0 | The baseline already refuses a bare "done"; kept as a regression guard | | execution-mode-routing | 0.50 | 0.50 | 0 | A mode is named in every run, but not a model per seat; the open gap in model-routing | | Mean | 0.83 | 0.63 | +0.20 | |

Measured on Claude Code 2.1.310 with sonnet as the agent model and haiku as the judge on 2026-10-25, cost $2.97 for the thirty sessions; raw results in docs/verification/. Run it yourself:

claude plugin eval . --trust-plugin --runs 3 --model sonnet --judge-model haiku --no-publish --max-cost-usd 12

Verified before every release

• node scripts/validate.js --ci: frontmatter rules, section order, cross-references, tool-name coupling, hook quoting and executable bits, version agreement across every manifest, README skill coverage, icon size.
• bash tests/test-hooks.sh on Linux, macOS and Windows Git Bash, and tests/test-hooks-windows.ps1 through cmd.exe.
• claude plugin validate . and agy plugin validate ..
• A headless acceptance session per verified harness: the prompt "Let's make a react todo list" must trigger rationale or intent-discovery before any code.
• bash tests/acceptance/claude-code-stop-check.sh: invoking completion-gate registers its stop check, the check evaluates the final message, and it removes itself after one pass.
• One review of the release diff by a different model family than the author.

What GodMode is not

• Not a code generator. It governs how the agent works, not what it builds.
• Not magic. Composable discipline in Markdown; read any skill and you know exactly what it asks for.
• Not free of judgement. Skills decide what evidence is needed; you still decide what to build.
• Not locked in. MIT. Every skill is a standalone file. Fork it, trim it, or run npx skills add and take only the ones you want.
• Not a telemetry source. No network calls, no analytics, no phone-home of any kind.

Contributing

Issues and pull requests are welcome. Read AGENTS.md first: it has the rules for skill wording (actions, not tools; facts from current sources; the bootstrap size limit) and the checks CI runs. New skills follow writing-skills.

Thanks

• @ivnad95, @g7trader and @AdversaLLC for diagnosing the 1.0 hook bug with exact reproductions; your fix is in 2.0.
• @urbabby, whose orchestrator-core fork skill shaped the "rulings, not stalls" idea in delegated-execution.
• Jesse Vincent and the superpowers contributors for the MIT-licensed Windows hook wrapper and multi-harness manifest layout that GodMode reuses.

License

MIT. See LICENSE.