| name | brain-dump-workflow |
| description | MANDATORY quality workflow for Brain Dump tickets. Defines the exact MCP tool call sequence every ticket must follow. Load this before starting any ticket work.
|
Brain Dump Universal Quality Workflow
Canonical source: src/api/ralph.ts → getRalphPrompt().
This skill provides a quick reference. The system prompt has the full workflow.
MANDATORY 5-Step Sequence
Every ticket MUST go through these steps using MCP tools. Never skip any.
Step 1: Start Work
Call workflow "start-work"({ ticketId: "<id>" }) BEFORE writing any code.
This creates a git branch, sets status to in_progress, and posts a "Starting work" comment.
Step 2: Implement + Verify
Write code, then discover and run this project's validation commands:
- Check project docs/config first:
AGENTS.md, CLAUDE.md, README, CONTRIBUTING, package scripts, pyproject.toml, go.mod, Makefile/Justfile, and CI files.
- Use the project's own commands. Do not assume pnpm, npm, TypeScript, lint, or test scripts exist.
- Common examples only: package script check/test/lint, pytest/ruff when configured,
go test ./..., cargo test, dotnet test, mvn test, ./gradlew test.
- If no automated validation command is discoverable, run a targeted manual smoke check and record that no project validation command was found.
- When available,
skill/brain-dump-workflow/scripts/run-quality-checks.sh can be used as a discovery helper.
Commit with format: feat(<ticket-short-id>): <description>
Step 3: Complete Implementation
Call workflow "complete-work"({ ticketId: "<id>", summary: "<what you did>" })
This moves ticket to ai_review and posts a work summary comment.
Step 4: AI Review (via MCP tools — NOT local review skills)
IMPORTANT: Do NOT use local /review skills, subagents, or code review tools.
Perform self-review by reading your own diffs, then record findings via MCP.
For each issue found, you MUST invoke the MCP tool:
review "submit-finding"({ ticketId, agent: "code-reviewer", severity, category, description })
Verify you receive a findingId in the response for each submission.
Fix critical/major issues, then invoke:
review "mark-fixed"({ findingId, fixStatus: "fixed" })
Verify: review "check-complete"({ ticketId }) — response must contain canProceedToHumanReview: true
Step 5: Generate Demo + STOP
Call review "generate-demo"({ ticketId, steps: [{ order, description, expectedOutcome, type }] })
Include 3-7 manual test steps. Ticket moves to human_review.
STOP HERE. Do NOT continue. Only humans can approve tickets.
DO NOT
- Skip any step above
- Set ticket status to "done" directly
- Continue working after generating demo
- Write code before calling
workflow "start-work"
- Use local review skills or subagents instead of
review "submit-finding"
- Describe demo steps in text instead of calling
review "generate-demo"
- Create git branches manually instead of using
workflow "start-work"
- Use any local alternative when an MCP tool exists for the step
Severity Guide (for Step 4)
| Severity | When to Use |
|---|
critical | Bug that breaks functionality or causes crashes |
major | Incorrect behavior or error handling issue |
minor | Code quality issue, not a bug |
suggestion | Nice-to-have improvement |
Demo Step Types (for Step 5)
| Type | When to Use |
|---|
manual | User performs an action |
visual | User visually confirms something |
automated | System runs a command/test |
Reference Docs
For detailed guidance on specific phases, see:
reference/review-guide.md -- Self-review checklist and review agent details
reference/troubleshooting.md -- Common errors and recovery steps