| name | forge-os |
| description | The Forge \u2014 cornerstone generation suite. Raw intent \u2192 production-grade prompt / workflow / skill / agent / plugin through one front door (/forge), five lanes, and a shared spine (translate \u2192 ground \u2192 compose \u2192 forge \u2192 gate \u2192 prove \u2192 wire \u2192 evolve). Model-dialect adaptive and platform-portable. Extends the existing generators; never rebuilds them. |
| version | 1.0.0 |
Forge OS — Raw Intent → Production-Grade Artifact
The system already forges world-class artifacts FROM artifacts (source → skill, prompt → skill,
extraction → skill). The Forge closes the one true gap: bare concept in, production-grade
artifact out — with the same deterministic rigor, plus two properties no single existing door
owns: (1) artifacts are born instrumented for self-evolution, and (2) artifacts are
model- and platform-agnostic by contract, adapted through a thin dialect layer instead of
per-artifact rewrites.
The Five Lanes
| Lane | Input | Output | Engine | Status |
|---|
| Prompt Forge | raw concept | one born-v2 structure-pure prompt + fixture | references/prompts-v2/prompt-forge.md | LIVE (Wave 1) |
| Workflow Forge | raw concept / repeated manual loop | .agent/workflows/<name>.md command with gates + verification | references/prompts-v2/workflow-forge.md | LIVE (Wave 2) |
| Skill Forge | raw concept (no source required) | full skill (SKILL.md + genius + workflows + v2 prompts) | Stage 1: references/prompts-v2/grounding-sprint.md (receipts-backed corpus) → existing /extract-forge pipeline | Stage 1 LIVE (Wave 2) |
| Agent Forge | raw concept or existing skill | agents/<name>/ AGENT.md + memory + card + /[name] front-door command (via sync_registries + generate_slash_commands) | references/prompts-v2/agent-forge.md (promotion mechanics per /create-agent; no-corpus → SKILL-FIRST stop) | LIVE (Wave 2) |
| Plugin Forge | skill / prompt-set / workflow family | installable plugin (plugin-dev toolchain) | references/prompts-v2/plugin-forge.md — HARD-GATED: builds nothing without a verbatim operator lift token + 4/4 lift-plan fixtures in one run | ENGINE READY · packaging gated (Wave 3) |
Routing honesty (never rebuild): if the intent arrives WITH an artifact, the Forge front door
routes to the existing crown jewels — source material → /extract-forge or /source-to-skill-system;
existing prompt → /convert-prompt; finished extraction → /convert-extraction; perfected exemplar
→ /design-skill-enshrine. The Forge's own lanes fire only for bare-intent input or for the
instrument/portability passes no other door owns.
The Shared Spine (every lane, in order)
- F0 TRANSLATE — Translation Card per
/raw-intent-bridge Stage 0: anchor, deliverable,
audience, felt standard VERBATIM, sharpened intent line. Felt-standard words travel to the
forging model untouched; the sharpened line does the routing.
- F1 GROUND — the Grounding Gate. Search prompt_library + skills registry + Recall + episodic
memory for owning expertise. Found → forge from that corpus (provenance guard). Not found →
grounding sprint (deep-research with receipts) manufactures an honest corpus FIRST, or the lane
flags
fidelity: low and forges less. Forging from training memory is prohibited
(docs/solutions/2026-07-07-transcript-only-extraction-generic-output.md).
- F2 COMPOSE — expert composition per the 5-slot standard (Spine / Differentiator / Mechanism /
Craft / Risk Gate; docs/solutions/expert-composition-standard.md). One owner, bounded experts.
- F3 FORGE — draft per the owning spec: prompts →
directives/prompt-forging-spec.md
(HIGH FLOOR, UNLIMITED CEILING); workflows → command conventions + gates; skills → skill-system
contract; agents → AGENT.md template; plugins → plugin-dev structure.
- F4 GATE — the deterministic layer:
renaissance_audit.py (prompts), verifier scripts,
prose/slop checks where prose ships. Existence ≠ done; audit-pass = done.
- F5 PROVE — live cold-start proof before registration: run the artifact on a real fixture in
a fresh context. Skills use the baseline-vs-with-skill differential; blind-pass against the best
in-library artifact of the same deliverable class backs any "top-1%" claim with a comparison it
could lose.
- F6 WIRE — registration: prompt_library build + wire_prompt_pointers + SLASH_COMMANDS.md +
invocation cards. An unwired artifact does not exist.
- F7 EVOLVE (born instrumented) — the artifact ships with 2–3 golden fixtures (input →
expected output shape) and a failure-logging pointer, wiring it into the EXISTING evolution
loop from birth:
/skill-anneal (≥3 documented failures), /self-evolve (plateaued),
evolution_orchestrator.py daily cycle. The Forge builds no new evolution engine.
Self-Evolution Spine — the model-relearn tax, deleted
Artifacts stay model-agnostic by contract: the v2 Output Contract + Skeleton + Quality Gate
enforce shape regardless of model. Model-specific behavior lives in ONE place:
- Model Dialect Cards —
directives/model-dialects/<model>.md, grown from
directives/model-notes.md. Quirks, structured-output behavior, param rules, verbosity tells,
instruction-following idioms.
/forge dialect <model> (LIVE, Wave 2) — the standard P1–P8 probe battery
(references/prompts-v2/dialect-probe.md) fingerprints a new model in one run and writes its
card. New model ships → one card updates → the entire forged library adapts.- Compliance replay — grown from
/test-model-compliance: on model change, replay each
artifact's golden fixtures, diff against expected shape, queue drift into the anneal loop.
Deterministic backstop, never AI-memory-dependent observability (standing rule).
Portability Spine — never trapped in one harness
The Platform Portability OS already owns cross-platform truth (per-platform constitutions +
execution/platform_compiler.py drift gate). The Forge compiles TO it: every forged artifact
keeps its methodology/contract core harness-neutral (plain markdown, bracket inputs, output
contracts) and isolates harness-specific wiring (hooks, frontmatter, tool names) in the thin
outer layer the platform compiler already manages. Result: prompts run anywhere; skills port by
constitution, not by hand-copying.
Boundaries
- No global writes (
~/.codex, ~/.claude) without explicit ask. No plugin marketplace edits.
- Plugin lane stays deferred until Farrice lifts it explicitly and cold-start proof passes.
- Extractions are never gated (standing decision 2026-06-09) — the Forge inherits that: forge
telemetry is telemetry, never a block.
- The Forge serves the whole system, unbounded by any single revenue goal (compass, never cage).
Source Spine (provenance for this skill's own prompts)
directives/prompt-forging-spec.md · .agent/workflows/raw-intent-bridge.md ·
.agent/workflows/source-to-skill-system.md · docs/solutions/expert-composition-standard.md ·
.agent/workflows/self-evolve.md · .agent/workflows/test-model-compliance.md ·
directives/model-notes.md. Nothing in this skill is forged from training memory.
Verification
After changing this skill or its prompts: python3 execution/renaissance_audit.py (0 fail) →
python3 execution/prompt_library.py build → python3 execution/wire_prompt_pointers.py --write.
Execution Prompts (structure-pure v2)
7 deterministic practitioner prompts — each carries an Output Contract, Output Skeleton, and Quality Gate. When a deliverable matches one, Read it and honor its contract instead of improvising the output shape.
- Agent Forge — Concept or Skill →
agents/<name>/ Deployable Agent — skills/forge-os/references/prompts-v2/agent-forge.md
- Dialect Probe — New Model → One Dialect Card —
skills/forge-os/references/prompts-v2/dialect-probe.md
- Grounding Sprint — Skill Forge Stage 1 —
skills/forge-os/references/prompts-v2/grounding-sprint.md
- Forge Front Door — Translation Card + Lane Decision —
skills/forge-os/references/prompts-v2/intent-translation-card.md
- Plugin Forge — Proven Asset Family → Installable Plugin —
skills/forge-os/references/prompts-v2/plugin-forge.md
- Prompt Forge — Bare Concept → Structure-Pure v2 Prompt —
skills/forge-os/references/prompts-v2/prompt-forge.md
- Workflow Forge — Bare Concept / Manual Loop →
.agent/workflows/<name>.md — skills/forge-os/references/prompts-v2/workflow-forge.md
Bundled Execution Prompt
Read ${CLAUDE_PLUGIN_ROOT}/skills/forge-os/references/prompt-forge.md and run it as written before producing output.
