/vibe — Unified Build Pipeline

Overview

You are orchestrating a 9-stage pipeline that takes an idea from validation to deployed and monitored MVP. Each stage invokes existing skills. State is tracked in VIBE-STATE.md at the project root.

Language: Match the user's language.

Startup Protocol

When /vibe is invoked:

1. Check for VIBE-STATE.md in the current working directory.
2. If it exists: Read it. Parse the stages table. Identify current stage and status.
3. If it does NOT exist: This is a fresh project. Initialize state as all pending.
4. Cross-check artifacts as backup confirmation:

• VALIDATION-REPORT.md → stage 1 done
• docs/superpowers/specs/*-design.md → stage 2 done
• docs/PRD.md → stage 3 done
• docs/superpowers/plans/*.md → stage 4 done
• HARNESS-STATE.md with COMPLETE or user-stopped → stage 5 done
• QA report or design review commits → stage 6 done
• Roast report PDF → stage 7 done
• Merged PR → stage 8 done
• Canary running → stage 9 active

1. If VIBE-STATE.md disagrees with artifacts, trust the artifacts and update VIBE-STATE.md.

Menu

Present this menu to the user:

/vibe

Pipeline: <project-name or "New Project">
Status: <current state summary>

  1. Validate        [<status>]
  2. Brainstorm       [<status>]
  3. Vibe Prep        [<status>]
  4. Plan             [<status>]
  5. Build            [<status>]
  6. QA + Design      [<status>]
  7. Roast            [<status>]
  8. Ship             [<status>]
  9. Monitor          [<status>]

Status key: [done] [in progress] [pending]

> Enter stage number, or press Enter to <suggested action>:

Suggested action logic:

• If a stage is in progress: "continue stage N"
• If all prior stages done and next is pending: "start stage N"
• If fresh project: "start from stage 1"
• User can always type any number to jump to that stage.

When user selects a stage, update VIBE-STATE.md to mark it in_progress, then execute that stage's process.

VIBE-STATE.md Management

Creating (fresh project)

When no VIBE-STATE.md exists, create it at the project root:

# Vibe Pipeline State

## Project: <ask user for project name>
## Started: <today's date>

## Stages
| Stage | Status | Completed | Notes |
|-------|--------|-----------|-------|
| 1. Validate | pending | — | — |
| 2. Brainstorm | pending | — | — |
| 3. Vibe Prep | pending | — | — |
| 4. Plan | pending | — | — |
| 5. Build | pending | — | — |
| 6. QA + Design | pending | — | — |
| 7. Roast | pending | — | — |
| 8. Ship | pending | — | — |
| 9. Monitor | pending | — | — |

## Roast Loops: 0

## Current Stage: —
## Next Suggested: Start stage 1

Updating

After each stage transition:

1. Read current VIBE-STATE.md
2. Update the relevant row's Status and Completed date
3. Update "Current Stage" and "Next Suggested"
4. If roast loops back to build, increment "Roast Loops" counter
5. Write the updated file

Stage statuses

• pending — not started
• in_progress — currently active
• completed — finished
• skipped — user chose to skip

Stage 1: Validate

Purpose: Kill bad ideas early by merging YC-style forcing questions with structured validation.

Skills used: gstack-office-hours + validate-idea

Process

1. Check for existing VALIDATION-REPORT.md.

• If exists: read it, show summary, ask: "Use this, re-validate, or skip?"
• If user says use it: mark stage 1 completed, go to menu.
• If user says re-validate: continue below.

1. Run office hours forcing questions.

Invoke the gstack-office-hours skill using the Skill tool in startup/builder mode. This covers:

• Demand reality: who is desperate for this?
• Status quo: what do they do today?
• Desperate specificity: what exact moment triggers the need?
• Narrowest wedge: what's the smallest version that solves it?
• Observation: what have you seen firsthand?
• Future-fit: does this get more relevant over time?

1. Run structured validation.

Invoke the validate-idea skill using the Skill tool. This covers:

• Socratic questioning
• First principles analysis
• Business model stress test
• Risk scanner (legal, tech, platform)
• Multi-role review panel (CEO, engineer, designer, skeptic)

1. Produce unified VALIDATION-REPORT.md.

Combine outputs from both into a single report at the project root. Include: GO / PIVOT / NO-GO verdict.

1. Gate: If verdict is NO-GO, warn the user strongly. They can override but the warning is on record.

1. Update VIBE-STATE.md: Mark stage 1 as completed.

1. Return to menu with stage 2 suggested.

Stage 2: Brainstorm

Purpose: Turn the validated idea into a design spec with multiple approaches explored.

Skills used: superpowers:brainstorming

Process

1. Read VALIDATION-REPORT.md for context on the validated idea.

1. Invoke `superpowers:brainstorming` using the Skill tool.

This will:

• Explore project context
• Ask clarifying questions one at a time
• Propose 2-3 approaches with trade-offs
• Present design in sections with approval per section
• Write design doc to docs/superpowers/specs/YYYY-MM-DD-<topic>-design.md
• Run spec self-review
• Get user approval on written spec

1. Update VIBE-STATE.md: Mark stage 2 as completed. Add note with spec file path.

1. Return to menu with stage 3 suggested.

Note: The brainstorming skill normally transitions to writing-plans. In this pipeline, we return to the menu instead so stage 3 (Vibe Prep) runs before stage 4 (Plan). When invoking brainstorming, after the spec is written and approved, do NOT let it invoke writing-plans. Return to the /vibe menu.

Stage 3: Vibe Prep

Purpose: Create PRD, UI design, architecture docs, and scaffold the project.

Skills used: vibe-prep (modified flow)

Process

1. Invoke `vibe-prep` using the Skill tool.

1. IMPORTANT: Skip vibe-prep's Step 1 (Idea Validation). Tell vibe-prep that validation is already complete by pointing it to the existing VALIDATION-REPORT.md. Start from vibe-prep's Step 2 (Tech Stack Selection).

1. Vibe-prep will handle:

• Tech stack selection (step 2)
• PRD writing — docs/PRD.md (step 3)
• UI/UX design — docs/UI-DESIGN.md (step 4)
• Architecture — docs/ARCHITECTURE.md (step 5)
• Project scaffold (step 6)
• Context documents — CLAUDE.md (step 7)
• Development standards — docs/REFERENCE.md (step 8)
• Git init & quality gates — docs/QUALITY-GATES.md (step 9)

1. Do NOT follow vibe-prep's handoff to vibe-harness. Return to the /vibe menu instead.

1. Update VIBE-STATE.md: Mark stage 3 as completed.

1. Return to menu with stage 4 suggested.

Stage 4: Plan

Purpose: Create a detailed, bite-sized implementation plan with TDD steps.

Skills used: superpowers:writing-plans

Process

1. Invoke `superpowers:writing-plans` using the Skill tool.

Provide context:

• Design spec from stage 2: docs/superpowers/specs/*-design.md
• PRD from stage 3: docs/PRD.md
• Architecture from stage 3: docs/ARCHITECTURE.md
• Quality gates from stage 3: docs/QUALITY-GATES.md

1. The writing-plans skill will:

• Scope check (decompose if needed)
• Map file structure
• Define bite-sized tasks with TDD steps
• Self-review against spec
• Save to docs/superpowers/plans/YYYY-MM-DD-<feature>.md

1. Do NOT follow writing-plans' execution handoff. The plan will be consumed by the enhanced vibe-harness in stage 5, not by subagent-driven-development or executing-plans directly.

1. Update VIBE-STATE.md: Mark stage 4 as completed. Add note with plan file path.

1. Return to menu with stage 5 suggested.

Stage 5: Build (Enhanced Vibe Harness)

Purpose: Autonomous 15+ cycle build loop with TDD and code review injected every cycle.

Skills used: vibe-harness + superpowers:test-driven-development + superpowers:requesting-code-review

Process

1. Invoke `vibe-harness` using the Skill tool.

1. CRITICAL ENHANCEMENT: Modify the Generator phase of EVERY cycle.

The standard vibe-harness Generator writes code freely. In this pipeline, the Generator MUST follow TDD discipline:

Enhanced Generator Phase (per cycle): a. Read the implementation plan from stage 4 to identify the next task/feature. b. RED: Write a failing test for the feature first. Run it. Confirm it fails. c. GREEN: Write the minimal implementation to make the test pass. Run it. Confirm it passes. d. REFACTOR: Clean up while keeping tests green. e. Commit the code.

Code Review Phase (per cycle, after Generator): f. Dispatch a code-reviewer subagent using superpowers:requesting-code-review. g. The reviewer checks the cycle's diff for quality, security, and spec compliance. h. If reviewer raises CRITICAL or HIGH issues: Generator fixes them before the Evaluator runs. i. If only MEDIUM or LOW issues: note them, proceed to Evaluator.

Evaluator Phase: Unchanged from standard vibe-harness. Scores on 5 dimensions.

1. All other vibe-harness behavior remains the same:

• 15 cycle minimum, COMPLETE banned before cycle 15
• HARNESS-STATE.md for persistence
• FEATURES.json tracking
• Dashboard at localhost:9999
• Living docs updated per cycle

1. On harness completion (COMPLETE verdict, user stop, or kill metric):

• Update VIBE-STATE.md: Mark stage 5 as completed.
• Return to menu with stage 6 suggested.

Stage 6: QA + Design Review

Purpose: Browser-based QA testing and visual polish of the built MVP.

Skills used: gstack-qa + gstack-design-review

Process

1. Start the dev server. Check package.json or project docs for the start command.

1. Run QA testing.

Invoke gstack-qa using the Skill tool. This will:

• Systematically test the web application in a headless browser
• Find bugs (broken links, form errors, console errors, responsive issues)
• Fix bugs iteratively with atomic commits
• Re-verify each fix with before/after screenshots

1. Run design review.

Invoke gstack-design-review using the Skill tool. This will:

• Find visual inconsistencies, spacing issues, hierarchy problems
• Detect AI slop patterns (generic gradients, over-rounded corners, etc.)
• Fix issues in source code with atomic commits
• Re-verify with before/after screenshots

1. Update VIBE-STATE.md: Mark stage 6 as completed.

1. Return to menu with stage 7 suggested.

Stage 7: Roast MVP

Purpose: Stress-test with AI personas. Auto-loop back to build if score is low.

Skills used: roast-mvp

Process

1. Invoke `roast-mvp` using the Skill tool.

Default to gauntlet mode (VC gate → community gate).

1. roast-mvp will:

• Run VC panel (5 personas) — must pass to unlock community
• Run community testing (10-100 personas)
• Generate PDF report with scores
• Collect user feedback on critiques
• Evolve persona gene pool

1. Auto-loop gate:

• If overall roast score < 6/10: Automatically return to Stage 5 (Build).
• Increment "Roast Loops" counter in VIBE-STATE.md.
• Pass roast feedback as context to the next vibe-harness run.
• The harness will use the roast critiques to guide its next cycles.
• Mark stage 7 as looped_back, stage 5 as in_progress.
• If overall roast score >= 6/10: Proceed.
• Mark stage 7 as completed.

1. Return to menu with stage 8 suggested (or stage 5 if looping back).

Stage 8: Ship

Purpose: Create PR and deploy to production.

Skills used: gstack-ship + gstack-land-and-deploy

Process

1. Run ship workflow.

Invoke gstack-ship using the Skill tool. This will:

• Detect and merge base branch
• Run tests
• Review diff
• Bump VERSION
• Update CHANGELOG
• Commit, push, create PR

1. Ask user to confirm deploy.

Show the PR link. Ask: "PR created. Ready to deploy? (y/n)"

• If no: mark stage 8 as completed (PR created but not deployed). User can deploy later.
• If yes: continue.

1. Run deploy workflow.

Invoke gstack-land-and-deploy using the Skill tool. This will:

• Merge the PR
• Wait for CI and deploy
• Verify production health via canary checks
• If deploy platform not configured: invoke gstack-setup-deploy first.

1. Update VIBE-STATE.md: Mark stage 8 as completed. Add note with PR URL and deploy status.

1. Return to menu with stage 9 suggested.

Stage 9: Monitor

Purpose: Continuous post-deploy health monitoring until manually stopped.

Skills used: gstack-canary + /loop

Process

1. Launch continuous canary monitoring.

Invoke gstack-canary using the Skill tool, wrapped in a /loop for continuous execution.

Use the Skill tool to invoke loop with args: 5m gstack-canary

This will:

• Run canary checks every 5 minutes
• Watch for: console errors, performance regressions, page failures
• Take periodic screenshots compared against pre-deploy baselines
• Alert on anomalies

1. Inform the user:

"Canary monitoring is running continuously every 5 minutes. You'll be alerted if any issues are detected. To stop monitoring, use /loop stop."

1. Update VIBE-STATE.md: Mark stage 9 as active (not completed — it runs until stopped).

1. Pipeline complete. Show summary:

   /vibe pipeline complete!
   
   Project: <name>
   Stages completed: 9/9
   Roast loops: <N>
   Final roast score: <score>
   PR: <url>
   Deploy: <status>
   Monitoring: active (stop with /loop stop)

Key Principles

1. Orchestrate, don't reimplement. This skill invokes existing skills. Do not duplicate their logic. If a skill's behavior needs to change (e.g., vibe-prep skipping validation), instruct the deviation explicitly but let the skill do its job.

1. VIBE-STATE.md is source of truth. Always read it before acting. Always update it after acting. If it disagrees with artifacts on disk, trust the artifacts and fix the state file.

1. Return to menu after every stage. Never auto-chain stages without showing the menu. The user must always have the option to jump, skip, or stop.

1. Respect skill boundaries. When invoking brainstorming, do NOT let it auto-transition to writing-plans. When invoking vibe-prep, do NOT let it auto-transition to vibe-harness. This skill controls the flow.

1. Roast auto-loop is mandatory. If roast score < 6, loop back to build. Do not ask. Do not skip. The user can override by jumping to stage 8 from the menu, but the default behavior is to loop.

Anti-Patterns

1. Skipping stages without asking. Always show the menu. Even if artifacts exist, confirm with the user.
2. Running stages out of order silently. If a user jumps to stage 5 without a plan (stage 4), warn them.
3. Ignoring VIBE-STATE.md. Every stage transition must be reflected in the state file.
4. Letting sub-skills chain to their default next step. This skill is the orchestrator. Sub-skills return here.
5. Losing roast feedback. When looping from stage 7 back to stage 5, the roast critiques MUST be passed as context to the harness. Otherwise the loop is pointless.