AI Project Workflow is a universal, AI-native software delivery workflow for Cursor, Claude Code, Codex, Qoder, CodeBuddy, TRAE, and other AI coding assistants. It keeps the project workflow in a single canonical core/ source and generates lightweight adapters for different tools.
AI coding agents can move too quickly when requirements, architecture, design, and tests are unclear. This workflow makes each stage explicit, writes deliverables to files, tracks state, and requires honest verification.
- Eight-stage workflow from initialization to retrospective.
- Core-first rules, Skills, Agents, templates, and schema.
- Platform adapters generated from the canonical
core/source. - Safe install behavior that does not overwrite user files unless
--forceis used. - Pure Node.js CLI with no Python requirement.
init -> prd -> hld -> sdd -> impl -> review -> deploy -> retro
Document priority:
PRD -> Architecture -> SDD -> Test Plan -> Code
- Cursor
- Claude Code
- Codex
- Qoder
- CodeBuddy
- TRAE
Use directly with npm:
npx @dayahs/ai-project-workflow init . --platform cursorOr install globally:
npm install -g @dayahs/ai-project-workflownpx @dayahs/ai-project-workflow init . --platform cursor
npx @dayahs/ai-project-workflow status .
npx @dayahs/ai-project-workflow validate .With a global install:
apw init . --platform cursor
apw status .
apw validate .apw init [target]
apw install [target]
apw sync [target]
apw validate [target]
apw status [target]
apw migrate [target]
apw --help
apw --versionSupported platforms:
cursor, trae, qoder, codebuddy, claude-code, codex, all
By default, the CLI does not overwrite user files. Conflicts are written as .ai-sdd.new files. Use --force only when overwrite is intended. Use --dry-run to preview writes.
- Cursor: uses
.cursor/rules/ai-sdd.mdcwith generated Skills and Agents. - Claude Code: uses
CLAUDE.mdas a lightweight entry point. - Codex: primarily reads
AGENTS.mdandcore/skills/. - Qoder: uses
.qoder/skills/and.qoder/agents/. - CodeBuddy: uses a minimal compatibility adapter and falls back to
AGENTS.md. - TRAE: preserves the slash command convention for the eight stages.
AGENTS.md
CLAUDE.md
VERSION
bin/
src/
core/
agents/
rules/
schemas/
skills/
templates/
adapters/
examples/
test-node/
core/ is the only canonical source. Adapter files should be generated, not manually maintained.
Target projects use .ai-workflow/state.json to track the current stage, completed stages, document status, skipped stages, blockers, and update time.
The schema is located at:
core/schemas/workflow-state.schema.json
apw validate .Validation checks required structure, Skill frontmatter, templates, schema JSON, Agent Skill references, residual terms, and adapter synchronization.
apw sync . --platform allSync regenerates platform adapters from core/.
apw migrate .
apw migrate . --applyMigration is dry-run by default. With --apply, legacy target-project layouts are moved into a backup folder inside the target project.
- Add
core/skills/<name>/SKILL.mdwith valid YAML frontmatter. - Use the standard sections: Goal, Use Cases, Preconditions, Required Context, Inputs, Allowed Changes, Steps, Outputs, Acceptance Criteria, Stop Conditions, Rollback Rules, Completion Report.
- If the Skill is part of the core stage flow, update
src/lib/constants.jsand tests. - Run
apw sync . --platform all. - Run
apw validate ..
- Add generation rules in
src/lib/adapters.js. - Keep
core/as the canonical source. - Mark generated files with
AUTO-GENERATED. - Add or update tests.
- Run sync and validation.
See examples/minimal-project/ for a minimal target project.
- Node.js >= 20
- npm
- No Python runtime is required
- The package provides workflow files and validation logic; it does not replace project-specific engineering judgment.
- Deployment steps depend on the target project's stack and environment.
- Security contact information should be configured by repository maintainers before public issue handling is relied on.
See CONTRIBUTING.md.
MIT. See LICENSE.