This repository is an AI-assisted product development system.
The system simulates a full product organization where specialized agents collaborate to build products.
The human product manager orchestrates the workflow using commands.
Agents represent specialized roles:
Research Agent Product Agent Design Agent Backend Architect Database Architect Frontend Engineer Backend Engineer Code Review Agent Peer Review Agent QA Agent Deploy Agent Analytics Agent Docs Agent Deslop Agent Metric Plan Agent Learning Agent
Each agent must only perform its assigned responsibility.
The system follows this workflow.
1 Capture Idea
/create-issue
Convert raw idea into structured opportunity.
Agent involved: Research Agent
2 Explore Opportunity
/explore
Validate the problem and analyze market.
Agent involved: Research Agent
3 Create Product Plan
/create-plan
Generate product specification, UX design, system architecture, and database schema.
Agents involved: Product Agent Design Agent Backend Architect Database Architect
4 Execute Plan
/execute-plan
Implement frontend and backend.
Agents involved: Frontend Engineer Backend Engineer
5 Deslop
/deslop
Clean and polish AI-generated code before review. Remove unnecessary complexity, fix naming, improve readability.
Agent involved: Deslop Agent
6 Code Review
/review
Review implementation quality.
Agent involved: Code Review Agent
7 Peer Review
/peer-review
Perform adversarial architecture review. Runs Challenge Mode: Assumption Audit, Anti-Sycophancy mandate, Multi-Perspective Challenge (Reliability Engineer / Adversarial User / Future Maintainer stances), and Prompt Autopsy Check. Optionally uses multi-model review — Claude leads and adjudicates, GPT-4o for bug analysis, Gemini for UI critique.
Agent involved: Peer Review Agent
8 QA Testing
/qa-test
Test system reliability.
Agent involved: QA Agent
9 Metric Plan
/metric-plan
Define post-launch metrics tracking, events, funnels, and success criteria.
Agent involved: Analytics Agent
10 Deployment Check
/deploy-check
Verify production readiness.
Agent involved: Deploy Agent
Blocking gates (all must pass before deploy approval):
- Build verification
- Environment configuration
- README quality gate (knowledge/readme-template.md standard)
- Sentry error tracking verification (@sentry/nextjs configured)
- Automated PR creation via
gh pr create(if all gates pass)
11 Postmortem
/postmortem
Analyze results and identify what went wrong across the full cycle. Includes mandatory Prompt Autopsy section: traces failures back to specific agent prompt instructions and proposes exact fixes to those agent files.
Agent involved: Learning Agent
12 Learning
/learning
Convert postmortem insights into durable system intelligence. Writes lessons to knowledge files, applies Prompt Autopsy findings directly to agent files, and generates apps/<project_name>/CODEBASE-CONTEXT.md for future agent sessions.
Agent involved: Learning Agent
These commands are not part of the sequential pipeline and have no quality gate dependencies. They can be run at any stage.
/docs Generate AI-native codebase documentation. Writes CODEBASE-CONTEXT.md to the active app directory (apps/<project_name>/CODEBASE-CONTEXT.md). Use after any significant implementation or architecture change, or before handing off to a new agent session. Agent involved: Docs Agent (or active engineering agent)
/explain Targeted learning session for the PM. Explains a concept, pattern, or error via the 80/20 rule tailored to a technical PM's knowledge level. Does not modify code or update pipeline state. Agent involved: none (Claude Code direct)
/eval Assertion-based grading of a completed issue's pipeline output against its spec. Run after /learning to verify that extracted insights improved the next cycle. Scores each assertion as PASS / FAIL / SKIP. Grade = PASS / (PASS + FAIL) * 100. Thresholds: 90–100% Excellent, 75–89% Good, 60–74% Needs improvement, <60% Critical. Does not modify code or update pipeline state. Agent involved: none (Claude Code direct)
Always follow the command workflow.
Do not skip validation steps.
Prefer small MVP experiments.
Prioritize user value and learning speed.
The human product manager is responsible for:
deciding which ideas to pursue evaluating agent outputs approving releases making final product decisions
Agents assist execution but do not replace judgment.
The AI Product OS enforces stage progression using quality gates.
A stage cannot proceed unless the previous stage passes its gate.
create-issue explore create-plan execute-plan deslop review peer-review qa-test metric-plan deploy-check postmortem learning
execute-plan cannot start unless: create-plan is complete
deslop cannot start unless: execute-plan is complete
review cannot start unless: deslop is complete
peer-review cannot start unless: review passes
qa-test cannot start unless: peer-review passes
metric-plan cannot start unless: qa-test passes
deploy-check cannot start unless: metric-plan is complete
postmortem cannot start unless: deploy-check passes
learning cannot start unless: postmortem is complete
Before executing a command:
1 Read project-state.md 2 Verify previous stage status 3 If gate not satisfied → stop execution 4 Return reason for failure
Before executing any command, the system must load the active project context.
Step 1 Read project-state.md.
Extract:
- project_name
- active_issue
- current_stage
Step 2 Load the issue file:
experiments/ideas/<active_issue>.md
Step 3 If available, load related documents:
experiments/exploration/exploration-<issue_number>.md experiments/plans/plan-<issue_number>.md
Step 4 Load knowledge base.
All agents must read:
knowledge/product-principles.md knowledge/coding-standards.md knowledge/architecture-guide.md knowledge/ui-standards.md knowledge/analytics-framework.md knowledge/prompt-library.md knowledge/engineering-lessons.md knowledge/product-lessons.md knowledge/ai-model-guide.md
For engineering commands (execute-plan, deslop, review, peer-review, qa-test, docs), also load if available:
apps/<project_name>/CODEBASE-CONTEXT.md
Step 5 Provide this context to all agents.
Agents must use this context when generating outputs.
Templates must not contain hard-coded product examples.
The AI Product OS determines the next command based on the current stage.
Pipeline Order
create-issue explore create-plan execute-plan deslop review peer-review qa-test metric-plan deploy-check postmortem learning
Resolution Logic
If current_stage = create-issue next_command = /explore
If current_stage = explore next_command = /create-plan
If current_stage = create-plan next_command = /execute-plan
If current_stage = execute-plan next_command = /deslop
If current_stage = deslop next_command = /review
If current_stage = review next_command = /peer-review
If current_stage = peer-review next_command = /qa-test
If current_stage = qa-test next_command = /metric-plan
If current_stage = metric-plan next_command = /deploy-check
If current_stage = deploy-check next_command = /postmortem
If current_stage = postmortem next_command = /learning
If current_stage = learning next_command = none — cycle complete. Run /create-issue to start next project.
After completing any command, the system must display the next recommended command to the user using the resolution logic above. If the stage is blocked, display the blocker instead.