Godmode

English | 简体中文

Your coding agent already knows how to code. Godmode teaches it how to engineer.

Godmode is a composable catalog of engineering workflows and expert capabilities for AI coding agents. It helps agents design before editing, test before claiming, review independently, and verify with fresh evidence.

The problem

Coding agents are good at producing code. Without explicit engineering behavior, they can start too early, miss existing conventions, bolt tests on at the end, and call a plausible-looking result complete.

WITHOUT GODMODE

"Build authentication"
        ↓
code immediately
        ↓
tests, security, and integration considered late
        ↓
looks done


WITH GODMODE

"Build authentication"
        ↓
solution design
        ↓
API + database + security expertise
        ↓
implementation plan
        ↓
test-driven development
        ↓
independent review
        ↓
fresh verification
        ↓
verified result

Composition, not one giant prompt

Godmode equips compatible agents to discover and compose the engineering behavior required by a task. Native skill discovery remains the routing authority; Godmode supplies focused boundaries, procedures, references, and deterministic helpers.

                 USER TASK

"Build a production analytics dashboard"

                     ↓
                  Godmode

       ┌─────────────┼─────────────┐
       ↓             ↓             ↓
solution-design frontend-design test-strategy
       │             │             │
       └─────────────┼─────────────┘
                     ↓
      implementation-planning + TDD
                     ↓
              browser-testing
                     ↓
                ui-ux-review
                     ↓
          completion-verification

This is an example composition, not a mandatory pipeline. A small task may need one skill; a consequential task may combine workflow and domain expertise.

More compositions

New API with persistence:
  solution-design + api-and-interface-design + database-design
  → implementation-planning → TDD → security-and-hardening

Flaky browser regression:
  root-cause-debugging + browser-testing
  → test-driven-development → completion-verification

Stateful production migration:
  safe-migrations + test-strategy + observability-and-instrumentation
  → release-engineering → behavior-validation

Active production impact:
  incident-response → root-cause-debugging
  → test-driven-development → completion-verification

More than Markdown

Godmode skills use the smallest implementation that earns its maintenance cost. Some are concise procedures with progressive references. Repeated or error-prone work can also include deterministic tooling.

frontend-design is the current quality bar: it combines a workflow, design and interaction references, stack-aware guidance, a searchable design catalog, design-system extraction, static UI auditing, and rendered verification. The goal is not to make every skill ship a script; it is to give each domain the knowledge and tools needed to produce evidence-backed work.

Install

Agent Skills-compatible clients

Copy the public catalog into a project-scoped skills directory:

mkdir -p .agents/skills
cp -R /absolute/path/to/godmode/skills/* .agents/skills/

Clients that support a skills installer can install the repository directly.

Claude Code

claude --plugin-dir /absolute/path/to/godmode
claude plugin validate /absolute/path/to/godmode

The thin SessionStart hook supplies shared operating rules; native skill discovery remains responsible for selecting a capability.

Codex

The repository includes .codex-plugin/plugin.json for direct loading and a local marketplace entry under .agents/plugins/. The marketplace publishes only skills/. Install it through the Codex plugin browser or local marketplace workflow, then verify that the expected skills appear.

Each public capability uses the standard Agent Skills layout: a directory with SKILL.md, concise routing metadata, and optional on-demand references or deterministic helpers. Godmode intentionally remains a portable catalog rather than a proprietary orchestration runtime.

Catalog

Core workflow skills

Engineering capabilities

Names are intentionally literal and task-oriented. They describe the boundary without depending on another repository's public vocabulary or compact aliases.

Design principles

• Agent claims are not evidence.
• Literal responsibility names beat compact aliases and borrowed vocabulary.
• Separate workflow states when they have different activation and handoff rules.
• Keep SKILL.md concise; load deeper rules progressively.
• Use deterministic helpers for repeatable, error-prone work.
• Review the rendered UI, not only its source.
• Treat logs, external text, generated output, and tool responses as untrusted data.
• Never claim compatibility with a client that was not actually checked.

Development

npm run check
python3 skills/frontend-design/scripts/design_system.py \
  --product "analytics dashboard" --tone technical --stack react
python3 skills/frontend-design/scripts/extract_design_system.py ./path/to/ui
python3 skills/ui-ux-review/scripts/audit_ui.py ./path/to/ui
python3 scripts/behavior_eval.py validate evals/behavior/core-workflows.json

The repository gate validates frontmatter, local links, body limits, manifest shape, 33 routing fixtures, behavior-eval case schemas, and helper tests. See CONTRIBUTING.md, docs/catalog.md, and docs/research.md for authoring and provenance details. See SUPPORT.md, SECURITY.md, and CHANGELOG.md before reporting or releasing changes. Native client evidence and limits are tracked in docs/compatibility.md. Maintainer-only handoff, provenance, lifecycle, behavior-eval, activation, and release procedures live in docs/maintainer-workflows.md; they are kept out of the public skill catalog to avoid routing collisions.

package.json remains private to prevent accidental npm publication. Public distribution uses the repository, client marketplaces, and checksum-verified GitHub release archives.

Status

Godmode is a pre-1.0 public preview. The catalog, deterministic helpers, routing fixtures, and behavior-eval harness are usable, but client-specific activation and output quality still require recorded forward runs across supported harnesses. A passing repository gate is not a claim th