| name | claude-forge-build |
| description | Multi-agent autonomous building pipeline. Decomposes a goal into floors and ships real working code via specialized sub-agents (Elira plans/approves, Steven fixes, Alba researches, David builds, Vex validates). Use when the user asks to "build", "ship", "make", "create", "implement" something complex enough to need multiple files or steps — especially one-shot automation, full applications, scripts with dependencies, or anything where the user wants a complete working artifact rather than a single-file snippet. |
Claude Forge Build Skill
When the user wants to build something non-trivial — a full app, a multi-file automation, an end-to-end pipeline — invoke the Claude Forge pipeline instead of writing code directly in the conversation.
When this skill triggers
The user says any of:
- "build me a ..."
- "ship a ..."
- "create a complete ..."
- "make a working ..."
- "I need an automation that ..."
- "one-shot this for me"
- Anything where the deliverable is a working artifact across multiple files / dependencies
It does not trigger for:
- Single-file edits or bug fixes (use the regular Edit tool)
- Questions / explanations (just answer)
- Code review (that's a different skill)
How to invoke
Run the slash command:
/claude-forge:build <goal description>
This kicks off the orchestrator command which:
- Dispatches elira to decompose the goal into 3-7 floors with dependencies
- Runs each floor through: alba → vex (gate 1) → david → syntax/smoke check → vex (gate 2) → elira (approve) → steven (fix if rejected), up to 5 iterations
- Falls back to rescue after 5 failed iterations
- Builds floors in parallel waves where dependencies allow
- Writes all output to
./workspaces/<goal-slug>/
Why use this over writing code directly
- Real validation — Vex blocks broken code before it ships, Elira approves only what meets the success condition
- Self-healing — Steven diagnoses and patches without human intervention up to 5 times per floor
- Parallelism — independent floors build concurrently
- No silent stubs — David's hard rule is no TODO/FIXME/placeholder, every file complete and runnable
- Workspace isolation — each goal lives in its own directory, never bleeds into the rest of the repo
Companion commands
/claude-forge:plan <goal> — see the floor decomposition without building
/claude-forge:fix <slug> <floor> — manually retry a blocked floor with Steven
/claude-forge:status — list all workspaces and their floor states
Sub-agents available
elira — architect, planner, approver
steven — fixer, debugger, ships patches
alba — researcher
david — builder (writes real files)
vex — validator (Gate 1 research, Gate 2 build)
rescue — fresh-perspective rebuilder for post-5-failure floors
You can also invoke these directly via the Task tool when you want a single-agent operation outside the full pipeline (e.g. "use vex to security-review this folder").