feat(skills): add content-culture calibration + use-real-subject

[vanceingalls]

Summary

• Content-culture calibration table in beat-direction.md — maps 6 genres to energy registers with explicit "wants" and "does NOT want" columns
• Use-real-subject rule in video-composition.md — use real artifacts when rights are clear, ask when unclear, abstract when unavailable

Changes

beat-direction.md

• Added "Content Culture Calibration" section before Per-Beat Direction
• 6 genres: consumer launch, product demo, technical explainer, documentary/archival, editorial/brand story, news/social
• Each genre has energy register, what it wants, what it does NOT want
• "Choose register FIRST, then enrich" — prevents the always-enrich over-rotation

video-composition.md

• Added "Use the Real Subject" section
• Three tiers: use when rights clear (public domain, user's own brand), ask when unclear (competitor brands, copyrighted text), abstract when unavailable
• Conservative default for commercial cases per review feedback

Review items addressed

• R4 reference removed (not present in rebased version)
• Stack rebased onto current feat(skills): design.md integration, shared video references, Claude Design gaps #549 tip
• Real-subject IP defaults softened — "ask when unclear" instead of "use by default"
• design.md schema gap resolved by feat(skills): design.md integration, shared video references, Claude Design gaps #549

Test plan

• Genre table consulted during Chevron, Ray-Ban, Friskies, and Starship picker tests this session
• Real-subject rule applied correctly during HeyGen brand extraction (used actual pacific codebase tokens)

🤖 Generated with Claude Code

[vanceingalls]

• feat(skills): pre-plan persistent-subject choreography (Addition C) #551
• feat(skills): add content-culture calibration + use-real-subject #550 👈 (View in Graphite)
• feat(skills): design.md integration, shared video references, Claude Design gaps #549
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[jrusso1020]

Staff review

TL;DR: The genre/energy-register table is genuinely good — it codifies a real failure mode ("treating a brand launch as introspective narrative") into a checkable rule. The "use the real subject" rule is also right in spirit. But there's a dangling cross-reference to R4 that isn't defined anywhere in the stack, and the real-subject rule glosses over copyright/trademark issues that matter for commercial output. Stacks on #549, so it inherits those issues too.

---

Blocking

1. R4 (persistent-element continuity morph) is referenced but never defined. I greppped the cumulative state of the stack (top of pr-551) — R4 appears only as a reference, never a definition. If this PR introduces the term, define it inline. If it's defined in #551, then this PR can't merge ahead of #551 without breaking the cross-ref. Given Graphite stacks land bottom-up (this is the middle), that's a sequencing issue worth flagging.

2. Real-subject rule has IP landmines that need a softer default. The current text says "use what's real" with three narrow exceptions. For commercial output that's the wrong default:

• Public-figure speeches: "quote their actual words when available" — recent speeches are still copyrighted (50–95+ year terms). The example glosses over this.
• Real company UI/brand: "use them unless the prompt says otherwise" — using a competitor's logo or another brand's UI in a marketing video without permission is trademark/passing-off risk, not a default-on behavior. This needs to flip: only use when the user has explicit rights, or for clearly non-commercial illustrative use.
• Museum open-access works: usually fine for public-domain art (Vermeer, Lange) but check IIIF rights metadata — some institutions restrict reproductions even of PD originals.

Suggestion: re-frame as "when the user has rights or the work is clearly public domain (verified, not assumed), use the real subject. When unclear, ask." That keeps the spirit (don't ship [ PHOTOGRAPH ] placeholders for the Migrant Mother piece) without inviting trademark/copyright headaches.

---

Significant

3. Stacks on unresolved #549 issues. The genre table tells Claude to "check design.md's mood field" when genre is ambiguous — but design.md schema isn't defined in #549. Until that's pinned down, "mood field" is aspirational.

4. Empty PR template + no test plan. Same issue as #549. For a rule that re-frames how every expansion behaves, please show at least 2–3 sample expansions across different genres (brand launch, documentary, technical explainer) — they're the easiest way to verify the table actually steers the model.

---

Smaller things

• Genre table line lengths: the "What does NOT want" column has long lines that may render awkwardly in markdown. Consider one-bullet-per-anti-pattern instead of comma-separated lists.
• Energy-register column is good but consider whether "Editorial / brand story" + "Documentary / archival" deserve separate rows or could collapse. They share a lot of "wants" (warmth, restraint).
• Forward-reference to R4 could be replaced with an inline parenthetical until #551 lands: e.g., "(see persistent-overlay choreography in #551)".

---

What this PR gets right

• The genre table is the kind of structured guardrail skills should have. A 6-row table mapping genre → energy register → wants → does-not-want is exactly the format that converts vague "tone" intuitions into checkable rules. The "running-shoes treated as introspective narrative" example is concrete and memorable.
• Naming the failure mode "always-enrich has a known over-rotation" is honest and useful. Calling out the regression risk in a rule that is mostly right makes the rule stronger.
• "Use the real subject" rule is correct in spirit — abstracting Migrant Mother to [ PHOTOGRAPH ] is exactly the failure mode worth fixing. Just needs a more conservative default for the commercial cases.

---

Recommended path

Block on (1) — define R4 here or merge after #551 lands. Soften the real-subject defaults (2). De-block on (3) once #549 lands. Fill in the PR template + ship 2–3 example expansions across genres for verification.

[jrusso1020]

Staff re-review (round 2)

TL;DR: No changes to this PR's own commits since round 1, but the stack drifted — #550 is based on ad44c313, which is 19 commits behind the current #549 tip. After Vance's round-1 fixes to #549 (gate merge, step renumbering, design-picker.md, beat-direction-format expansion), this PR will conflict heavily on references/prompt-expansion.md and reintroduce some of the issues that #549 just fixed. Restack required before merge.

---

New blocker

0. Stack is stale. git merge-base origin/vance/04-28-prompt-expansion origin/vance/04-28-content-culture resolves to ad44c313, but #549's tip is 541173f8. Concretely:

• #549 rewrote references/prompt-expansion.md's "What to generate" into the new beat-direction format (Title + style block / Rhythm declaration / Global rules / Per-scene beats with concept-mood-depth-choreography-transition / Recurring motifs / Negative prompt).
• This PR's prompt-expansion.md still has the old 7-section format (Title / Global animation rules / Scene-by-scene breakdown with bg/mg/fg sub-bullets / Recurring motifs / Transition rules / Pacing curve / Negative prompt).
• Merging this PR as-is would revert the new beat-direction format that landed in #549.

Run gt restack (or gt sync --no-interactive) and resolve prompt-expansion.md keeping #549's new "What to generate" — then layer this PR's content-culture and use-real-subject sections on top.

---

Round 1 blockers — status (after restack)

1. R4 is referenced but never defined. Still BLOCKING. I re-grepped the full stack including #549's latest commits (541173f8):

$ git grep -nE "^- R[0-9]|^R[0-9]|## R[0-9]|R1 |R2 |R3 |R5 " origin/vance/04-28-persistent-choreography
(no results)

R4 has no definition — only parenthetical glosses. If this PR introduces the term "persistent-element continuity morph", define it as a numbered rule somewhere (or drop the R4 notation and call it out by name inline).

2. Real-subject IP defaults too aggressive. Unchanged. The "Real company UI/brand → use them unless the prompt says otherwise" default still has trademark/passing-off exposure. Recommended re-frame: "When the user has rights or the work is verifiably public domain, use the real subject. When unclear, ask." That keeps the spirit (don't ship [ PHOTOGRAPH ] placeholders for Migrant Mother) without inviting commercial-use risk for competitor logos.

3. design.md schema gap (carried from #549) → ✅ Resolved by #549 — picker now outputs google-labs-code/design.md spec. So when #550's content-culture text says "check design.md's mood field", the schema exists. Good.

4. PR template + test plan. Still empty. For a rule that re-frames how every expansion behaves, ship 2–3 sample expansions across genres (brand launch, documentary, technical explainer) — table-stakes verification.

---

What this PR continues to get right

• The genre/energy-register table is the strongest piece in the stack. Mapping genre → energy register → wants → does-not-want is exactly the format that converts vague "tone" intuitions into checkable rules. The "running-shoes treated as introspective narrative" example is concrete.
• Naming "always-enrich has a known over-rotation" as an honest regression risk is good defensive design.
• "Use the real subject" rule is correct in spirit — the failure mode (abstracting Migrant Mother to [ PHOTOGRAPH ]) is real and worth fixing.

---

Recommended path

1. Restack on current #549 tip (541173f8). Resolve prompt-expansion.md by keeping #549's new beat-direction "What to generate" and inserting this PR's content-culture + real-subject sections at the appropriate point.
2. Define R4 inline (or drop the rule-number notation).
3. Soften the real-subject defaults for commercial UI/brand cases.
4. Add 2–3 sample expansions to the PR description for genre verification.