The Forge

A portable methodology for high-quality work with a team of AI agents: unbounded recursive generations, each running divergence -> adversarial critique -> verification, then retrospect and handoff. It improves both the product and the process that builds it.

The spec is FORGE.md - vendor-neutral. Hand it to any capable agent (Claude Code, OpenAI Codex, Cursor, Gemini, or a single agent with no sub-agents) and it can run the system by mapping the phases onto its own primitives.

The handoff template is HANDOFF.md - the API between generations. A fresh agent resumes from it with no chat history, so a run can continue indefinitely across fresh contexts.

Use it with any agent

• Inline: paste or reference the spec in your prompt - "Run the Forge (see FORGE.md) on <task>."
• Repo-level: for agents that read a repo instructions file, point to it from there. In AGENTS.md (Codex) or CLAUDE.md (Claude Code):

  When asked to "run the Forge", follow ./FORGE.md.

Use it in Claude Code (skill)

Install via skills.sh:

npx skills add JEJ111/the-forge

This installs the self-contained skill from skills/forge/ (the runbook plus bundled FORGE.md and HANDOFF.md). Or install it as a personal skill manually by copying skills/forge/ to ~/.claude/skills/forge/.

Invoke with /forge, or just say "run the Forge on ...". The skill orchestrates it with the Workflow tool across budget-bounded generations and treats FORGE.md as the source of truth. (Restart Claude Code if the skill is not listed yet.)

Use it in OpenAI Codex (skill)

Installed as a personal skill at ~/.codex/skills/forge/. Codex auto-discovers it; invoke with $forge or just say "run the Forge on ...". The skill maps the phases onto Codex's primitives - parallel sub-tasks when the session can spawn them, otherwise a sequential run resumed across fresh contexts - and treats FORGE.md as the source of truth. (Restart Codex if the skill is not listed yet.)

What it is / is not

• Is: a process for confidence, craft, and momentum - build, have adversaries try to break it, prove it with a real gate, then hand off so a fresh agent continues. Unbounded generations with deterministic safety rails, bounded depth (Lead -> Primary -> Helper); it improves itself as it runs, and composes established patterns (Reflexion, evaluator-optimizer, the Ralph loop) rather than reinventing them. Scaled to the stakes.
• Is not: a product or a specific tool. The phases map onto whatever orchestration your agent has.

Sharing

This repo is self-contained. Other teams or agents only need FORGE.md and HANDOFF.md to run the methodology - drop them into any project, or reference them from your agent instructions file (AGENTS.md, CLAUDE.md). The Claude Code skill lives in skills/forge/ and is installable via npx skills add JEJ111/the-forge.

License

MIT © 2026 Jeremy Jackson. Use it, fork it, adapt it to your own agents.

---

Unbounded generations. Bounded depth. Fresh agents. Clean handoffs. Verified deltas. Deterministic guardrails. Self-improving process.