New 29th sub-skill higgsfield-facs — direct a face by muscle (FACS Action Unit codes like AU12 lip-corner puller, AU6 cheek raiser) instead of emotion labels. The platform's highest-resolution facial control: forced/uncanny/mixed expressions, micro-performance, and honest acting in close-up dialogue.

Provenance: model-checked against the live seedance_2_0 spec, which exposes no FACS/expression field — so the technique is flagged [EMPIRICAL] end to end. The AU vocabulary itself is standard, citeable science; Seedance's interpretation of codes in a prompt is the empirical, not-a-guarantee part.

Highlights

• Plan-first workflow — plan 3–4 expressions → generate a FACS sheet for only those → write the codes. Explicitly counters the "generate the full 49-AU sheet then cherry-pick" anti-pattern.
• FACS reference-sheet generation — parameterized image prompt (GPT Image 2 / Nano Banana Pro) + the LLM-mislabels-AUs caveat (the circulating sheet's AU82 vs standard AU38).
• Codes-only vs codes+anatomical-description, the 3–4-expressions-max accuracy limit, optional character photo.
• Full AU code table + EMFACS emotion→AU recipes (Duchenne AU6+AU12, sadness AU1+AU4+AU15, fear, anger, surprise, disgust, contempt).
• Dialogue & monologue facial acting — AU-per-beat schedule combined with the [AUDIO: Xs] lip-sync block; three worked examples.

Wiring + cross-links

• Routed from root SKILL.md, sub_skill_descriptions.py, INDEX.md; new templates/seedance/facs-expression-beats.md. Roster 28 → 29.
• Cross-linked from higgsfield-soul (Micro-Expressions), higgsfield-seedance (Voice Rewrite), higgsfield-audio (Lip-Sync Rules), and vocab.md (Emotion-as-Visible-Behavior channels).

Verification: validate.py --strict clean · 110 tests pass · 40/40 evals.