Fornace
usage ยท Fornace Pi
pi
๐Ÿค– Agent-Readable Endpoint: For autonomous models and scripts, fetch the raw markdown directly: curl -fsSL https://onboarding.fornace.app/usage.md view raw .md โ†—

Pi operator & agent manual.

Essential commands, model routing, media rules, and prompt workflows for the Fornace team.

Slash commands & operations

CommandTypeDescription
/fornace install setup Fresh provisioning of team skills (122), local extensions, and DramaFoundry credentials.
/fornace update update Idempotent reconciliation. Checks manifest sha256 digests, refreshes DramaFoundry tokens, updates skills.
/fornace list info Lists installed extensions, active skill count, and verification status.
/fornace doctor check Verifies team gateway key, Matchbox passkey state, DramaFoundry credentials, and tools.
/fast session context Mechanical context compaction (zero model calls). Fires on the first idle moment. Call when context approaches 50%.
/compact context Interactive model-driven summary compaction of recent turns.

Model routing & delegation

The Fornace Mantice gateway (https://llm.fornace.net/v1) serves cost-optimized model tiers. Always prefer the least expensive capable tier.

Tier / AliasSubagent RoleCapabilities & Intended Tasks
fornace-flash quick Ultra-low latency, cheap mechanical work, digests, RTK re-reading, small edits.
fornace-fast scout Fast reconnaissance, regex search, pattern matching, initial bug triage.
fornace-reasoning planner, researcher High-depth synthesis, architectural roadmaps, complex design review.
fornace-max builder, critic, operator Substantial coding, multi-file refactoring, production delivery, adversarial review.
fornace-astra recovery High-tier recovery. Explicit user request only. Never use for routine turns.

Generative media doctrine

RULE ONE No traditional deterministic VFX

Our code never assembles, blends, patches, warps, or composites picture content. Every pixel of a repaired, altered, or generated frame must come out of an AI model. Code only trims clips, prepares model inputs, calls APIs, and measures results.

RULE TWO Approved editors only, proposed first

Approved video editors: Wan 2.7 edit (fal-ai/wan/v2.7/edit-video) and FLUX 3 edit. Every edit must be proposed with the source status, defect, and intended change, and approved by human eye before running.

long-form-character-video

2 to 3 minute vertical story films. Fast pace (190-210 wpm), active physical business (props, gestures), populated public worlds, scene-specific sound design, and an adversarial still gate before full render.

skill ~/.pi/agent/skills/long-form-character-video

thinksound-native

Native C++17 + Metal 4 local sound design on Apple Silicon. Foley, ambience, and scene sound beds mixed under dialogue at minus 33 LUFS. Zero Python runtime.

metal 4 ~/.pi/agent/skills/thinksound-native

dynamic-subtitles

Word-by-word highlighted captions burned with Remotion and local mlx-whisper ASR word timestamps. Layout scales cleanly to vertical 9:16.

remotion ~/.pi/agent/skills/dynamic-subtitles

DramaFoundry production

Full-scale short drama series platform: manuscript ingest, episode director planning, character turnarounds, per-beat video generation, and sound mixing. Read the complete DramaFoundry guide โ†—.

Web UI: https://dramafoundry.fornace.net  |  Credentials: ~/.pi/agent/credentials/dramafoundry.json Guide โ†— Web UI โ†—

Paid advertising & growth suite

The 23-skill ads suite covers full-funnel reconnaissance, creative generation, and account audits:

/ads dna <url>  โžœ  /ads create  โžœ  /ads generate  โžœ  /ads audit

Copyable agent prompt templates

Audit this repository against our AGENTS.md rules. Run a freshness check on the active SDKs and platforms in play. Check that source files are โ‰ค 400 lines, no dead fallbacks exist, native Apple Silicon is prioritized, and report actionable findings.
Using the dramafoundry skill, inspect the screenplay at ./screenplay.txt. Generate the beat breakdown for Episode 1, define the character identity sheets for FLUX, and verify our model gateway credentials.
Run autonomous-qa against http://localhost:3000. Simulate a new user journey, test core forms and navigation, catch console errors or DOM blockers, and repair issues in place.