Prometheus Planning: Interview-Style Requirements Gathering and Work Plan Generation

What You'll Learn

• Start a Prometheus planning session to clarify project requirements through interview mode
• Understand Prometheus's core principle of "plan only, don't implement"
• Collaborate with Metis and Momus to generate high-quality, comprehensive work plans
• Use the /start-work command to hand off plans for Atlas execution

Your Current Challenge

Imagine giving AI a complex task: "Help me refactor the authentication system."

5 minutes later, AI starts writing code. You're happy, feeling you've saved time.

30 minutes later, you realize:

• AI didn't ask which auth library to use (JWT? NextAuth? Session?)
• AI made many assumptions (like "must support OAuth" when you don't need it)
• Code is halfway done, but direction is wrong, complete redo needed
• Testing reveals core logic incompatible with existing system

This is the classic "mixed planning and execution" problem: AI starts working before requirements are clear, leading to extensive rework.

When to Use Prometheus

When to Use

Suitable for Prometheus:

• Complex feature development (e.g., "add user authentication system")
• Large-scale refactoring (e.g., "refactor entire data access layer")
• Architecture design (e.g., "design microservices architecture")
• Tasks requiring strict quality assurance

Direct Sisyphus execution:

• Simple bug fixes (e.g., "fix login button typo")
• Clear, small features (e.g., "add form with 3 inputs")

🎒 Preparation

Ensure the following configurations are complete:

• [ ] Prometheus agent enabled (enabled by default)
• [ ] At least one AI Provider configured (Anthropic, OpenAI, etc.)
• [ ] Understanding of basic agent concepts (completed "AI Agents Team: Meet 10 Experts")

Verify Prometheus is available:

# Type in OpenCode chat
@prometheus

# Should see Prometheus prompt:
# "Hello, I am Prometheus - Strategic Planning Advisor. ..."

Core Concepts

Prometheus's Core Identity Constraint

What's Prometheus's most important feature? It never writes code.

Function	Prometheus	Sisyphus	Atlas
Requirements gathering	✅	❌	❌
Work plan generation	✅	❌	❌
Code implementation	❌	✅	✅ (delegated)
Task execution	❌	✅	✅ (delegated)

Why is this important?

• Planner ≠ Executor: Just like product managers don't write code, Prometheus's role is "how to do it," not "doing it"
• Prevent assumptions: If Prometheus could write code directly, it might "guess and do" when requirements are unclear
• Force thinking: After being banned from writing code, Prometheus must clarify all details

Three-Phase Workflow

flowchart LR
    A[User input requirements] --> B[Phase 1: Interview Mode]
    B -->|Requirements clear?| C[Phase 2: Plan Generation]
    C --> D[Metis consultation]
    D -->|Need high precision?| E[Momus loop verification]
    E -->|Plan refined| F[Generate .sisyphus/plans/*.md]
    C -->|No high precision needed| F
    F --> G[/start-work execution]

Phase responsibilities:

• Phase 1 - Interview Mode: Collect requirements, research codebase, continuously update draft
• Phase 2 - Plan Generation: Consult Metis, generate complete plan, present summary
• Phase 3 - Execution: Hand off to Atlas via /start-work

Follow Along

Step 1: Start Prometheus Planning Session

Why Trigger Prometheus via keyword or command to enter interview mode.

Action

Type in OpenCode chat:

@prometheus help me plan a user authentication system

You should see:

• Prometheus confirms entering interview mode
• Asks first question (e.g., "What tech stack is your app built with?")
• Creates draft file .sisyphus/drafts/user-auth.md

Key Feature: Draft Files

Prometheus continuously updates files under .sisyphus/drafts/. This is its "external memory":

• Records decisions from each discussion
• Saves discovered code patterns
• Marks clear boundaries (IN/OUT)

You can check the draft anytime to verify Prometheus's understanding is correct.

Step 2: Answer Questions, Let Prometheus Gather Context

Why Prometheus needs to "understand" your project to generate executable plans. It doesn't guess—it gains insight by researching codebase and best practices.

Action

Answer Prometheus's questions, for example:

User input:
My app is Next.js 14, App Router, currently no authentication.
I want to support email/password login and GitHub OAuth.

What Prometheus will do:

• Use explore agent to analyze existing code structure
• Use librarian agent to find authentication best practices
• Update "Requirements" and "Technical Decisions" sections in draft file

You should see:

I've started the explore agent to analyze your project structure...

1. explore: Find existing session patterns
2. librarian: Find NextAuth best practices

Waiting for results to return, then I'll continue with questions.

Step 3: Check Draft File (Optional)

Why The draft is Prometheus's "external memory." You can verify at any time if its understanding is correct.

Action

# Check draft content in terminal
cat .sisyphus/drafts/user-auth.md

You should see similar content:

# Draft: user-auth

## Requirements (confirmed)
- Tech stack: Next.js 14, App Router
- Auth methods: Email/password + GitHub OAuth
- Current state: No authentication implementation

## Technical Decisions
- No decisions yet

## Research Findings
- Explore agent running...

Step 4: Continue Answering Until Requirements Are Clear

Why Prometheus has a "Clearance Checklist." Only after all items are checked will it automatically transition to plan generation.

Prometheus's judgment criteria:

CLEARANCE CHECKLIST (ALL must be YES to auto-transition):
□ Are core objectives clear?
□ Are scope boundaries clear (IN/OUT)?
□ No critical ambiguities left?
□ Is technical solution determined?
□ Is testing strategy confirmed (TDD/manual)?
□ No blocking issues?

Action

Continue answering Prometheus's questions until it says:

All requirements are clear. Consulting Metis and generating plan...

You should see:

• Prometheus calls Metis agent
• Metis analyzes potential missed issues
• Prometheus adjusts understanding based on Metis's feedback

Step 5: Review Generated Plan

Why The plan file is Prometheus's final output, containing all tasks, dependencies, and acceptance criteria.

Action

# Check generated plan
cat .sisyphus/plans/user-auth.md

You should see complete structure:

# User Authentication System

## Context
[Original requirement description]
[Interview summary]
[Metis analysis results]

## Work Objectives
- Core objective: Implement email/password login and GitHub OAuth
- Specific deliverables: Login page, API endpoints, session management
- Completion definition: Users can log in and access protected routes

## Verification Strategy
- Infrastructure exists: YES
- User wants tests: TDD
- Framework: vitest

## TODOs
- [ ] 1. Install NextAuth.js and configure
  - References:
    - https://next-auth.js.org/getting-started/installation
  - Acceptance Criteria:
    - [ ] npm run test → PASS (1 test)

- [ ] 2. Create API route [...nextauth]/route.ts
  - References:
    - src/lib/session.ts:10-45 - Existing session pattern
  - Acceptance Criteria:
    - [ ] curl http://localhost:3000/api/auth/... → 200

- [ ] 3. Implement login page UI
  - References:
    - src/app/login/page.tsx - Existing login page structure
  - Acceptance Criteria:
    - [ ] Playwright verification: See login form
    - [ ] Screenshot saved to .sisyphus/evidence/

...

Step 6: Choose Execution Method

Why Prometheus gives you two choices: quick start or high-precision review.

Prometheus's presentation (using Question tool):

## Plan Generated: user-auth

**Key Decisions Made:**
- Use NextAuth.js (integrates well with Next.js App Router)
- GitHub OAuth provider + email/password

**Scope:**
- IN: Login functionality, session management, route protection
- OUT: Registration, password reset, user profile editing

**Guardrails Applied:**
- Must follow existing session pattern
- Do not modify core business logic

Plan saved to: `.sisyphus/plans/user-auth.md`

---

**Next Step**

How to proceed?

1. **Start Work**: Execute /start-work. The plan looks solid.
2. **High Accuracy Review**: Let Momus strictly verify every detail. Adds review cycles but guarantees precision.

Action

Select an option (click button in OpenCode or type option).

).

If you choose "Start Work":

• Prometheus deletes draft file
• Prompts you to run /start-work

If you choose "High Accuracy Review":

• Prometheus enters Momus loop
• Continuously fixes feedback until Momus says "OKAY"
• Then prompts you to run /start-work

Step 7: Execute Plan

Why The plan is Prometheus's output; execution is Atlas's responsibility.

Action

# Type in OpenCode
/start-work

You should see:

• Atlas reads .sisyphus/plans/user-auth.md
• Creates boulder.json status file
• Executes each TODO in sequence
• Delegates tasks to specialized agents (e.g., UI work to Frontend)

Checkpoint ✅

• boulder.json file created
• Atlas starts executing task 1
• Status updates after each task completion

Common Pitfalls

Pitfall 1: Rushing for Plan Before Requirements Clear

Problem:

User: @prometheus do a user authentication
User: Don't ask so many questions, just generate the plan

Consequence: Plan full of assumptions, requires repeated modifications during execution.

Correct approach:

User: @prometheus do a user authentication
Prometheus: What tech stack is your app? Any auth currently?
User: Next.js 14, App Router, no auth
Prometheus: What login methods do you need?
User: Email/password + GitHub OAuth
...
(Continue answering until Prometheus auto-transitions)

Remember this principle

Planning time ≠ wasted time

• Spend 5 minutes clarifying requirements to save 2 hours of rework
• Prometheus's interview mode is saving your future self time

Pitfall 2: Not Checking Draft Files

Problem: Prometheus records many decisions and boundaries in draft, but you don't check before asking it to generate plan.

Consequences:

• Plan contains incorrect understanding
• You realize during execution "I never asked for this!"

Correct approach:

1. After starting planning, constantly monitor .sisyphus/drafts/ files
2. Correct misunderstandings immediately: "No, I don't want OAuth, just simple JWT"
3. Continue after correction

Pitfall 3: Splitting Plans into Multiple Generations

Problem:

User: This project is too big, let's plan phase 1 first

Consequences:

• Context fragmentation between phase 1 and phase 2
• Inconsistent architectural decisions
• Requirements missed across multiple sessions

Correct approach:

✅ Single Plan Principle: No matter how big, all TODOs go in one .sisyphus/plans/{name}.md

Why?

• Both Prometheus and Atlas can handle large plans
• Single plan ensures architectural consistency
• Avoids context fragmentation

Pitfall 4: Forgetting Metis's Role

Problem:

User: Requirements are done, quickly generate the plan
Prometheus: (Generates directly, skipping Metis)

Consequences:

• Plan might miss key boundaries
• No "Must NOT Have" explicitly excludes scope
• AI slop appears during execution (over-engineering)

Correct approach:

✅ Metis consultation is mandatory, no need to rush it

What does Metis do?

• Identifies questions Prometheus should have asked but didn't
• Raises boundaries that need clarification
• Prevents AI over-engineering

Pitfall 5: Ignoring Testing Strategy Decisions

Problem: Prometheus asks "Do you need tests?" and you say "doesn't matter" or skip.

Consequences:

• If testing infrastructure exists but TDD not utilized, opportunity missed
• If no tests and no detailed manual verification steps, execution failure rate is high

Correct approach:

Prometheus: I see you have the vitest testing framework. Should the work include tests?
User: YES (TDD)

Impact:

• Prometheus structures each task as: RED → GREEN → REFACTOR
• TODO Acceptance Criteria will explicitly include test commands
• Atlas will follow TDD workflow during execution

Lesson Summary

Prometheus Core Value:

• Separate planning from execution: Through "no code writing" constraint, ensures clear requirements
• Interview mode: Continuous questioning, researching codebase, updating drafts
• Quality assurance: Metis consultation, Momus validation, single plan principle

Typical workflow:

1. Input @prometheus [requirement] to start planning
2. Answer questions, check .sisyphus/drafts/ draft
3. Wait for Prometheus auto-transition (Clearance Checklist all checked)
4. Review generated .sisyphus/plans/{name}.md
5. Choose "Start Work" or "High Accuracy Review"
6. Run /start-work to hand off to Atlas for execution

Best practices:

• Spend time understanding requirements, don't rush for the plan
• Continuously check draft files, correct misunderstandings promptly
• Follow single plan principle, don't split large tasks
• Clarify testing strategy, affects entire plan structure

Next Lesson Preview

Next, we'll learn Background Tasks: Working Like a Team.

You'll learn:

• How to make multiple agents work in parallel, greatly improving efficiency
• Configure concurrency limits to avoid API rate limiting
• Manage background tasks, retrieve results and cancel operations
• Coordinate multiple expert agents like a "real team"

---

Appendix: Source Code Reference

Click to expand source code locations

Last Updated: 2026-01-26

Function	File Path	Line Number
Prometheus system prompt	src/agents/prometheus-prompt.ts	19-1184
Prometheus permission config	src/agents/prometheus-prompt.ts	1187-1197
Metis agent	src/agents/metis.ts	-
Momus agent	src/agents/momus.ts	-
Orchestration guide	docs/orchestration-guide.md	67-90

Core Constants:

• PROMETHEUS_SYSTEM_PROMPT: Complete system prompt defining Prometheus's identity, workflow, and constraints

Key Functions/Tools:

• PROMETHEUS_PERMISSION: Defines Prometheus's tool permissions (only allows .md file editing)

Business Rules:

• Prometheus default mode: INTERVIEW MODE
• Auto-transition condition: Clearance Checklist all items are YES
• Metis consultation: Mandatory, executed before plan generation
• Momus loop: Optional high-precision mode, loops until "OKAY"
• Single plan principle: No matter how large the task, all TODOs in one .md file
• Draft management: Continuously updates .sisyphus/drafts/{name}.md, deletes after plan completion