Created
March 19, 2026 01:24
-
-
Save dreamiurg/4ee601c1140c82ea2bf511573b8399c2 to your computer and use it in GitHub Desktop.
War Room: AI-coordinated engineering team for real-time bug fixing and design
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| # War Room: AI-Coordinated Engineering Sessions | |
| ## What It Is | |
| The `/war-room` skill turns Claude Code into a **team lead** that coordinates a full engineering team of AI agents in real-time. The human tests the product, reports bugs, and makes decisions — Claude orchestrates everything else. | |
| ## How It Works | |
| The human starts a war room session, and Claude: | |
| 1. **Spawns a team** of specialist agents (backend engineers, frontend engineers, architect, QA, SRE, designers, Linear/Notion experts, merge queue manager) | |
| 2. **Picks up tickets** from Linear, creates tracking beads, assigns work to engineers | |
| 3. **Runs a conveyor belt**: fix → quality review → PR → CI monitoring → merge | |
| 4. **Monitors autonomously** with recurring loops: | |
| - Linear sync (every 5 min) — catches tickets claimed by other engineers | |
| - SRE log patrol (every 15 min) — scans DataDog for new errors | |
| - Notion transcript patrol (every 15 min) — extracts bugs/ideas from live meeting transcripts | |
| - Work pickup (every 5 min) — assigns idle engineers to unblocked tickets | |
| - Skill reflection (every 30 min) — reviews friction points and improves its own process | |
| ## What Happened in Today's Session (5 hours) | |
| ### Engineering Output | |
| - **12 bug fixes** shipped and merged (one PR, 16 commits) | |
| - **2 BLOCKERs caught** by automated code review before they reached production: | |
| - Sync cycle silently overwriting user-edited descriptions with null | |
| - Frontend API wrapper checking for HTTP 404 after backend changed to 200+null | |
| - **1 proactive investigation** found a real bug: Inngest retry silently dropping media events | |
| - **1 feature built** (individual selection for import wizard) | |
| ### Design Output | |
| - **2 interactive HTML playgrounds** with multiple layout variants: | |
| - Appointment header redesign (4-group info architecture) | |
| - Client detail page (unified timeline, sticky memory sidebar, real data) | |
| - **2 UX briefs** with data inventory, interaction models, edge cases | |
| - **1 comprehensive UI spec** (component tree, state matrix, migration plan) | |
| ### Process | |
| - **14 Linear tickets** created/updated with bead cross-references | |
| - **5 Notion transcript items** caught and filed automatically from the live call | |
| - **15+ protocol improvements** codified into the skill during the session | |
| - **0 manual reminders needed** — the autonomous loops kept work flowing | |
| ## Key Concepts | |
| ### The Conveyor Belt | |
| Every fix goes through: Engineer implements → Merge queue runs quality gates → `/simplify-until-converge` catches issues → PR created → `/babysit-pr` handles CI and code review → Merge. The human doesn't touch any of this. | |
| ### Linear as Coordination Layer | |
| Multiple humans with multiple Claude instances coordinate through Linear. Before claiming work, agents check if someone else already has it. Every ticket gets a bead ID comment for traceability. | |
| ### File Write Hold Protocol | |
| With 20+ agents in one worktree, git push can fail when agents write files during the pre-push hook. Solution: broadcast HOLD → all agents confirm → push → broadcast RESUME. | |
| ### Notion Transcript Patrol | |
| War rooms run alongside live calls with Notion AI transcription. The patrol scans the transcript every 15 minutes, files beads for bugs/ideas discussed on the call, and presents them to the human for approval before creating Linear tickets. | |
| ### Team Scaling | |
| The team scales based on available work. Started with 2 engineers, scaled to 4 when the backlog allowed parallelization, added UX experts and UI designers when design work was needed. | |
| ## The Team Composition | |
| | Role | What They Do | | |
| |------|-------------| | |
| | **Team Lead** (Claude) | Orchestrates, never writes code | | |
| | **Backend Engineer(s)** | NestJS, Prisma, Zoom API fixes | | |
| | **Frontend Engineer(s)** | React, shadcn/ui, state management | | |
| | **Architect** | Root cause analysis, design tradeoffs | | |
| | **QA** | Test checklists, edge cases | | |
| | **SRE** | DataDog monitoring, log analysis | | |
| | **Merge Queue** | Quality gates, pushes to remote | | |
| | **Simplify Agent** | Multi-round code review until convergence | | |
| | **Babysit-PR** | CI monitoring, CodeRabbit feedback, merge | | |
| | **Linear Expert** | Ticket operations, sync checks | | |
| | **Notion Expert** | Meeting notes, transcript patrol | | |
| | **UX Expert** | Information architecture, interaction design | | |
| | **UI Designer** | Interactive HTML playgrounds with variants | | |
| | **DB Engineer** | Read-only database queries, data analysis | | |
| ## The Skill File | |
| The war room skill is a single markdown file (~400 lines) that defines: | |
| - Startup sequence (analyze changes, create team, spawn teammates) | |
| - Operating protocol (triage, dispatch, verify) | |
| - Conveyor belt stages | |
| - Autonomous monitoring loops | |
| - Bug classification rules | |
| - File write hold protocol | |
| - Design team workflow | |
| - Session close checklist | |
| It improves itself every 30 minutes based on observed friction points. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment