Project Forge

Project Forge is an AI-agnostic Agent Skills package for turning a product or game concept into an executable project contract, repository scaffold, ownership plan, evidence harness, and independent review loop.

The package does not depend on a specific model, agent API, or sub-agent implementation. Agents with parallel delegation can fan out independent work. Agents without it execute the same implementer, critic, and verifier roles as isolated sequential passes.

Skills

• project-forge: discovers, contracts, scaffolds, and coordinates a project.
• project-forge-review: independently reviews an integrated project and records evidence-backed findings.

The first production template is threejs-game. Additional templates should be added only after they have their own fixtures, validation rules, and acceptance evidence.

Install

After the repository is published:

npx skills add moinsen-dev/project-forge --skill project-forge -a codex

Install both skills for several agents:

npx skills add moinsen-dev/project-forge \
  --skill project-forge \
  --skill project-forge-review \
  -a codex \
  -a claude-code \
  -a cursor

Use the bootstrap skill once without installing it:

npx skills use moinsen-dev/project-forge@project-forge --agent codex

For local development:

npx skills add . --list
npm run check

Validate directly against the Agent Skills reference implementation:

uvx --from skills-ref agentskills validate skills/project-forge
uvx --from skills-ref agentskills validate skills/project-forge-review

Initial Prompt

Use project-forge to bootstrap an original Three.js game from this concept:

A physics-driven cave rescue game with inertia, limited fuel, a tractor beam,
dangerous terrain, and short replayable missions.

Create the product contract, architecture, ownership plan, quality gates,
repository scaffold, and first playable milestone. Make explicit assumptions
where the concept is incomplete. Do not begin broad production work until the
discovery contract and first vertical slice are coherent.

Generated Contract

A threejs-game bootstrap produces:

project.yaml
prompt.md
ARCHITECTURE.md
AGENTS.md
ASSIGNMENTS.md
QUALITY_GATES.md
forge-state.json
evidence/

project.yaml contains JSON, which is valid YAML 1.2. This keeps the contract machine-readable without adding a runtime YAML dependency.

Development Principles

• Product rules and acceptance evidence are agent-independent.
• Capability differences change scheduling, not quality gates.
• Missing browser, visual, or performance evidence stays unverified.
• Implementers cannot approve their own work.
• Parallel work requires disjoint ownership.
• Coupled systems are integrated and tuned sequentially.
• References expose quality gaps; they are not permission to copy.
• A generated scaffold is the beginning of discovery, not proof of quality.

License

MIT