LumabyteCo/clarifyprompt-mcp

Summary

Patch release. Two additive ships, no engine code changes:

1. packs/higgsfield-creative-handbook.md — knowledge pack for Higgsfield-targeted prompts (Soul ID workflow, camera moves, model-selection rules, Marketing Studio modes, prompt structure). Pairs with the higgsfield platform entries from 1.6.1.
2. npm run matrix — multi-model eval matrix runner that produces a side-by-side HTML showing engine performance across model classes.

Higgsfield knowledge pack

9 H2-chunked sections covering:

• Model selection guide — which of the 13 Higgsfield-routed models for which use case (Soul 2.0 for portraits, Sora 2 for photoreal video, Cinema Studio for filmic shots, Kling 3.0 for motion fidelity, etc.)
• Prompt structure pattern — long-form natural-language prose with the canonical structure ([shot type] [subject + action] [setting] [lighting] [textures] [palette + mood] [style cues])
• Camera moves — explicit vocabulary table (push-in, pull-out, crane, orbit, tilt, pan, handheld, etc.) for video models
• Soul ID workflow — 3-step training process for face-faithful character reuse across generations
• Multi-reference editing — when to route to edit?model=multi instead of text-to-image
• Marketing Studio workflows — brand kit, avatars, products, hooks, virality predictor
• Mode presets — UGC Factory, TV Spot, Wild Card, Pinterest Pin, etc.
• Common pitfalls — don't translate Midjourney flags verbatim, don't redescribe Soul ID characters in prose, keep < 150 words
• Output specs — 4K images, 15s video, credit costs

Context Curator pulls relevant chunks via semantic retrieval when prompts target Higgsfield. No explicit loading needed for users who npm install clarifyprompt-mcp — the pack ships in the npm tarball and is discovered at runtime.

npm run matrix — multi-model eval matrix

npm run matrix -- --models llama3.2:3b,qwen2.5-coder:7b-instruct-q4_K_M,qwen2.5:14b-instruct-q4_K_M

Runs the eval harness sequentially against each model, writes per-model JSON, stitches into one side-by-side HTML at evals/matrix.html.

What it lights up:

• The 3 model-class-gated fixtures (shape-small-local-model / shape-mid-tier-model / shape-reasoning-model) that single-model runs skip — each lights up on its target model class.
• Cross-model deltas at a glance — e.g. analyzer-creative-media fails on qwen-coder-7b but passes on gpt-4o-mini.
• Hover any cell for failing-check details.

Implementation note: evals/run.mjs gained a --json-out <path> flag (matrix.mjs uses it; also useful for CI tooling that wants machine-readable run data).

Where it lives: evals/matrix.mjs is NOT in the npm tarball — it's a maintainer/contributor tool you run from a cloned repo, not a runtime artifact.

Compatibility

• No MCP tool surface changes — still 23 tools.
• No env-var changes.
• No platform count changes — still 60+ platforms.
• Tarball grows ~10 KB for the knowledge pack file (103 files / ~162 KB packed total).

Full diff: https://github.com/LumabyteCo/clarifyprompt-mcp/compare/v1.6.1...v1.6.2