← ClaudeAtlas

forgelisted

FORGE protocol — world-class multi-model agent swarm and prompt optimiser. Fable engineers a rubric-driven plan (true goal + success rubric + constraints + risks + minimal-sufficient MECE decomposition); Opus critiques & repairs it for MECE/minimality/critical-path; sub-tasks route by tier+kind with effort scaling and real tool use; a perspective-diverse voter panel (with a soundness veto) verifies; failures escalate UP a tier and re-verify; Fable synthesizes and grades the final answer against the rubric. Cost-aware routing: haiku → sonnet → opus → fable. Runs automatically on every message via the UserPromptSubmit hook, and can be invoked manually as /forge <prompt>.
somudatta06/forge · ★ 0 · AI & Automation · score 75
Install: claude install-skill somudatta06/forge
# FORGE The mind is **Fable**. Workers are **Haiku / Sonnet / Opus**. Nothing is wasted: cheap tasks go to Haiku, real reasoning to Sonnet/Opus, and only the two hardest jobs — *enhancing the prompt* and *synthesizing the answer* — go to Fable. This protocol runs automatically (via the hook) OR when invoked as `/forge <prompt>`. When invoked manually, treat the args as the raw prompt. --- ## The phases (v2.1 — world-class) Pipeline: **Enhance (Fable)** → **Critique+Repair plan (Opus→Fable)** → **Swarm (tier+kind routed, tool-using)** → **Verify (voter panel + soundness veto)** → **Escalating self-repair + re-verify** → **Synthesize + rubric-grade (Fable)**. The four roles, each made world-class: - **Prompter (Enhance):** extracts the *true goal*, writes an explicit *success rubric* (the 10/10 bar), surfaces implicit constraints + failure modes, then does a *minimal-sufficient* MECE decomposition (fewest units, no busywork, shortest critical path). - **Router:** each unit carries a **kind** (retrieval/computation/reasoning/generation/synthesis/verification) plus a tier; model + effort + tool-use follow from them. Failed units **escalate UP a tier** on repair (haiku→sonnet→opus). - **Optimiser (Critique):** Opus enforces MECE, minimality, correct tiering, and critical-path parallelism; Fable repairs the plan once. - **Doer:** workers actually run tools when `needsExecution`/`needsSearch` is set; a **soundness veto** forces repair on any honesty failure even if outvoted; r