| name | forge-autonomous-work |
| description | Run a bounded autonomous work loop that generates competing approaches, implements iteratively, runs acceptance checks, stress-tests the result, learns from the output, and either ships or reports why it cannot. Use when the user asks to "forge this", "run autonomously", "take this end to end", "iterate until done", "iterate until tests pass", or provides runnable acceptance criteria for hands-off implementation. Produces an artifact, evidence log, passing checks or escalation record, and final verdict. |
Forge Autonomous Work
Forge is a process engine for substantial work. It does not take the first plausible plan. It heats alternatives, hammers one into shape, quenches it under pressure, tempers it by measuring what happened, and anneals if the structure is brittle.
Use this for complex coding, research, writing, planning, or operational tasks where success can be checked.
0. Preflight
Before starting, define:
- Goal
- Scope
- Hard boundaries
- Budget or iteration cap
- Verification method
- Stop conditions
- Acceptance commands, if this is an implementation task
If there is no checkable end state, do not forge. Use a planning or dream skill first.
When the user provides a SPEC/GOAL/ACCEPTANCE block, run Forge in acceptance mode:
SPEC: short name
GOAL: observable outcome
ACCEPTANCE:
- runnable command or check
FILES: optional scope
MAX_ITERATIONS: optional, default 5
Stop and clarify if every acceptance criterion is manual or opinion-based, the goal is vague, or the allowed file scope is risky.
1. Heat: generate competing approaches
Create 2-3 meaningfully different approaches. For each:
- Strategy
- Files/artifacts affected
- Expected benefit
- Known risks
- Verification method
- Testable hypothesis
