gooseworks-ai/goose-skills / skills/video/video-production-harness
video-production-harness
VerifiedFull video production sequence with script review, actual ingredient choices, controlled generation, editing, evidence-based quality review, polish, captions and delivery. A host binding supplies project storage, authentic approvals, provider access and billing.
states/setup-and-script.md
View on GitHub---
name: video-production-harness/states/setup-and-script
description: States 0 to 2.6 of the shared video production sequence — preflight, brand context, reference study, brainstorm, design brief, script lock and voice selection.
---
# Human version
The first part of the run: check tools and inputs, load the brand, study the reference, pick the concept, write the brief and the script, and choose the voice. The script approval happens here.
---
# Agent version
Part of [the orchestrator](../orchestrator.md). Its non-negotiables, approvals and decision policies apply. Read one file per call; if the output says it was truncated, re-read it in parts before acting.
### State 0 — Pre-flight audit
1. Reuse a PASS audit under 24 h old only when tooling, host authentication and every input fingerprint are unchanged.
2. Otherwise follow [preflight-audit.md](../preflight-audit.md).
3. **Only a hard P0 finding stops the run:** broken media, unsupported formats, missing authentication or capabilities, and unresolved brief variables. Fix them before generation. No local override bypasses a gate.
4. P1 warnings (unknown product size, a creative-contract inconsistency) are logged in the state file with a severity ([check severity](../references/check-severity.md)) and don't block; an `ask` one goes to the customer at the script approval.
5. **Ask the review mode once.** If the customer hasn't said, `review_mode` is `ask`. When they say "just make it" or "don't ask me", record `auto` with their words through the host ([check severity](../references/check-severity.md)); money approvals still apply.
### State 0.1 — Brand bootstrap and reference study
1. Load brand_context and brand_assets through the host, not an assumed folder. Confirm identity, product and offer, audience, claims and evidence, visual identity and existing creative grammar.
2. If context is missing, offer one combined choice: research now with an available research capability, accept supplied facts and assets, continue with the stated limits, or defer. Never invent claims or treat an empty template as research. Paid research needs an estimate and the host's authorization. Save the decision so a resumed run does not ask again.
3. Reload brand context at the design, storyboard, world-lock and end-card stages. Pick assets by their labeled provenance, not by filename.
4. **Reference.** When there is one, wait for its prepared media and study it as [creative contract and review](../references/creative-contract-and-review.md) says: the whole moving picture and its actual sound, with the beat table (people, action, camera, sound, text, length). Record how you perceived it (`listened`/`watched` or `measured`). Frames plus a transcript are never called a full watch. Never invent analysis or use a reference as the final.
### State 1 — Brainstorm
1. If `idea_brief` and a schema-valid `concepts` entry exist, go to State 2.
2. Otherwise follow [brainstorm.md](../brainstorm.md). It writes `idea_brief` and `concepts`. Check both parse and carry every field the binding's schema needs. Save the full concept, title, format, hook, logline, evidence and risks on the review surface.
3. Keep an existing concept selection; carry open options into the script review in State 2.5.
4. **Concept-fit check:** *who is this selling, to whom?* If the answer is not "[brand] to [their real customer]", reframe now. Ask through the host decision surface when unclear.
### State 2 — Design brief
1. If `design_brief` and `implementation_brief` exist and match the current inputs, go to State 2.5.
2. Follow [create-design-brief.md](../create-design-brief.md). Its step 4 sets the scene skeleton and routes every scene through [shot routing](../references/shot-routing.md). Save the creative_contract (speech_mode, soundtrack_modes, required_features, reference_map, changes_from_reference, product_grounding, voice_plan), `climax_line:` and `assets_manifest:`; bind later work to the returned creative_contract_revision.
3. Map a legacy `audio_strategy` through [speech and route qualification](../references/speech-and-route-qualification.md) before branching. Native dialogue never becomes unrelated narration or silence.
4. No pause by default. If `<auto>` is false and real creative choices came up (style, voice options), show them in the open-decisions section at the script gate.
### State 2.5 — Script lock
1. Reuse `locked_script` only if its content and revision still match the host-approved script.
2. **Fetch write-video-ad-script with catalog_fetch and follow it to write the script**; it is a required skill of this playbook. If the fetch fails, stop and tell the customer: never write the script inline. Pass the sourced creative brief to its preparation command with `--brief working/script/creative-brief.json`; use its documented fields and set `shape.requires_creative_brief: true` for new custom or template work. Measure the reference with `scripts/reference_rate.py <reference words JSON> --target-s <seconds>`; never estimate the rate. Record `speaking_rate_wps` and its source file in the script plan's `voice_plan` and size the script to its `script_words` ([review-script-shortform](../references/review-script-shortform.md)).
3. Then follow [lock-script.md](../lock-script.md) with `<concept_type>` from the idea brief.
4. **SCRIPT OWNER GATE.** Show the complete words, the essential action and proof, the sound plan, the feasible routes and, in bold at the top, every `changes_from_reference` entry and the shot and setup counts next to the reference's (the `reference_shots.py compare` line). In the same message ask: "Do you want an on-camera person speaking, or voiceover over footage?", suggesting what the reference does; record it in `voice_plan.route` ([speech and route qualification](../references/speech-and-route-qualification.md)). The approval shows both `approval_quote.tier_quotes` choices, **Standard** (the default) and **Best**, for the same `draft.shot_list`, each with its total and `yes_covers_up_to` as given; a choice with `over_credit_cap` is never offered (shrink the plan and re-quote); record the owner's pick as `quality_tier` (sent to `video_approve` with that choice's total), and a shot they name at the other choice ("Best for the pouring shot") in `draft.shot_overrides` `[{scene, tier?, seedance?: true, user_quote}]`, saved and re-quoted before the yes. The tier sets each shot's starting setting ([shot routing](../references/shot-routing.md)). Record the approved revision; a later material contract change refreshes the affected approval.
5. After approval, `locked_script` is the only source of VO text for every later step.
### State 2.6 — Voice and performance selection
1. Read the resolved speech_mode, soundtrack_modes and delivery intent. No-speech work needs no audition; a song or native dialogue uses its selected performance.
2. Choose from [speech qualification](../references/speech-and-route-qualification.md) and [talking-actor production](../references/talking-actor.md). Exact supplied audio requires qualified audio-driven lip-sync; realistic speech needing a selected timbre may use qualified joint voice-and-picture generation. Do not replace a supported route or its quote silently. Narration and non-speaking scenes keep their own sound plan.
3. **Voice and character are one choice.** Match the voice to the on-screen character's gender, age, ethnicity, accent and energy and to the reference narrator's persona and pace. Never default to a voice used for another brand or character. Record the pair in `voice_plan`.
- **The person comes first.** For every generated on-screen person who speaks, run [lock-character](../lock-character.md) Phase 0 now, before any audition and even when no audition is needed: it makes the person's anchor candidates and, for a separate voice (voice first and lip-sync, or narration of that person), `character.voice_pair`. That pair is the voice; never pick it earlier from a name or the voice list, and never put a voice id in the script draft (`voices`) before it exists. The anchor is picked at the State 3 ingredient gate, and State 3.5 continues from it at Phase 1.
- **Native speech** (one unconditioned H3 take speaking its own line) has no separate voice: the person is made with `--no-voice-pair`, and the generated voice is reviewed with that take.
- When no library voice fits, prefer an ElevenLabs Voice Design voice made from the same character descriptors, when the host offers it. Where it can't run, `pair_voice.py --character <character.json> --accept-nearest` takes the closest `nearest` voice as the pair; the owner hears it beside the still, and a rejection re-pairs with `--exclude` as below. When no voice of the person's heritage is in the lists, `nearest` falls back to the closest voices of the right gender across heritage and says so (`differs_on` includes `heritage`). Voice Design stays the first choice, but the customer may take the nearest voice instead (free) after hearing it beside the still; tell them in one plain line, for example "No voice in the library matches her heritage, so this is the closest one." **Never change the person (heritage, age or gender) to fit a voice.**
- **Changing the voice later** re-runs `pair_voice.py --character <character.json> --voices <list> --exclude <rejected voice_id>` and uses the voice it returns. Never hand-pick a replacement from the list.
4. For new actor casting, show three candidates, choose the anchor, then play two or three compatible voices reading the same approved excerpt. Include a brand pronunciation, a conversational turn and the longest line where useful. Reuse an accepted actor/voice and valid sample; do not regenerate for a resumed run. Paid auditions use the existing preview budget, are priced before calls, and stay inside its approved bound. Pairing labels and automated scores do not choose the voice for the customer.
5. Register playable auditions with exact lines, beat ids, voice settings, script revision and provenance. Hear the isolated audio when your tools can; text, waveform or transcript cannot establish quality. Show open voice choices in the existing ingredient review beside the selected actor (narration needs no actor). Explicit creative delegation may select but still shows the chosen sample. Never present an unplayed audition as chosen. For a voice-reference route, retain its pair and compare the actual generated voice with that audition after ingredient approval; an unconditioned native route has no promised ElevenLabs voice.
6. Reuse approved audition audio where it covers a production line. Render only missing or changed lines with the selected settings after the gate, on `eleven_v4`. Do not force a louder climax. Measure the voice's pace as create-clips Phase 4 says.
7. Recheck every speech scene and the full mix in States 5, 8 and 10. On resume, restore the selected ids; a missing selection blocks only the route that needs it.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| receiving-code-reviewobra/superpowersUse when receiving code review feedback, before implementing suggestions, especially if feedback seems unclear or technically questionable - requires technical rigor and verification, not performative agreement or blind implementation | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| requesting-code-reviewobra/superpowersUse when completing tasks, implementing major features, or before merging to verify work meets requirements | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| using-git-worktreesobra/superpowersUse when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallback | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| verification-before-completionobra/superpowersUse when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |