Start branch: lab/05-start
Solution tag: lab/05b-solution
Session replay: labs/05b/session.jsonl
Objective
Execute the wave plan you created in Lab 05a. Watch agents implement your design in parallel, review their work, merge PRs, and end up with a working MCP server registered in your Claude Code session.
This lab teaches the execution phase of the workflow.
Prerequisites
Steps
Step 1: Verify your plan is ready
Do: Check that your /prepwaves plan from Lab 05a is still active.
Verify: Task list shows wave tasks with issue references
Learn: Plans survive between sessions via the task list.
Step 2: Execute Wave 1
Do: Run /nextwave. Watch planning agents, flight partitioning, and execution.
Verify: Wave 1 PRs are merged to main
Learn: The flight model prevents merge conflicts by detecting file-level overlaps before agents start coding.
Step 3: Execute remaining waves
Do: Run /nextwave for each remaining wave.
Verify: All waves complete, all PRs merged
Learn: Each wave builds on the previous one's merged code. This is why dependency order matters.
Step 4: Install your MCP server
Do: Register your server with Claude Code:
claude mcp add --scope project --transport stdio my-server -- bun mcp-server/index.ts
Verify: claude mcp list shows your server
Learn: MCP servers are registered per-project or per-user. --transport stdio is the simplest transport.
Step 5: Test it for real
Do: Ask Claude to use each of your server's tools. Verify the results.
Verify: All tools respond correctly when called by Claude
Learn: The ultimate test — not "does the test suite pass" but "does it actually work when Claude calls it?"
Step 6: Retrospective
Do: Review what happened: How closely did implementation match your PRD? Did any agent escalate? Were there merge conflicts?
Verify: You can articulate one thing you'd do differently in the design phase
Learn: Design quality determines execution quality. The retrospective closes the feedback loop.
You Learned
Congratulations
You've completed the full Claude Code workflow — from PRD to running code, with parallel agent execution, automated code review, and wave-based delivery. The MCP server you built is yours to keep and extend.
Stuck?
Load labs/05b/session.jsonl into Clawback to watch an example execution session.
Start branch:
lab/05-startSolution tag:
lab/05b-solutionSession replay:
labs/05b/session.jsonlObjective
Execute the wave plan you created in Lab 05a. Watch agents implement your design in parallel, review their work, merge PRs, and end up with a working MCP server registered in your Claude Code session.
This lab teaches the execution phase of the workflow.
Prerequisites
install.sh --checkall green)Steps
Step 1: Verify your plan is ready
Do: Check that your
/prepwavesplan from Lab 05a is still active.Verify: Task list shows wave tasks with issue references
Learn: Plans survive between sessions via the task list.
Step 2: Execute Wave 1
Do: Run
/nextwave. Watch planning agents, flight partitioning, and execution.Verify: Wave 1 PRs are merged to main
Learn: The flight model prevents merge conflicts by detecting file-level overlaps before agents start coding.
Step 3: Execute remaining waves
Do: Run
/nextwavefor each remaining wave.Verify: All waves complete, all PRs merged
Learn: Each wave builds on the previous one's merged code. This is why dependency order matters.
Step 4: Install your MCP server
Do: Register your server with Claude Code:
Verify:
claude mcp listshows your serverLearn: MCP servers are registered per-project or per-user.
--transport stdiois the simplest transport.Step 5: Test it for real
Do: Ask Claude to use each of your server's tools. Verify the results.
Verify: All tools respond correctly when called by Claude
Learn: The ultimate test — not "does the test suite pass" but "does it actually work when Claude calls it?"
Step 6: Retrospective
Do: Review what happened: How closely did implementation match your PRD? Did any agent escalate? Were there merge conflicts?
Verify: You can articulate one thing you'd do differently in the design phase
Learn: Design quality determines execution quality. The retrospective closes the feedback loop.
You Learned
Congratulations
You've completed the full Claude Code workflow — from PRD to running code, with parallel agent execution, automated code review, and wave-based delivery. The MCP server you built is yours to keep and extend.
Stuck?
Load
labs/05b/session.jsonlinto Clawback to watch an example execution session.