# Higgsfield production and cross-business playbook

Reviewed: 2026-09-07 local / 2026-09-08 UTC. Owner: Production Director. This supplements the Q3 integration strategy with live access findings, precise failure handling and reusable procedures.

## Verified now

Connected read-only balance, model catalog and sandbox checks succeeded. The visible private workspace is free with zero credits and no unlimited allowance. No generation benchmark has run. The catalog snapshot in `data/operations/higgsfield-capabilities.snapshot.json` captures five candidates from the first 30 video entries; it is not a complete inventory. Refresh before use. Current marketing calls Cinema Studio 4.0, while the inspected model ID is `cinematic_studio_3_0`; marketing names and executable IDs are distinct. [Supercomputer](https://higgsfield.ai/supercomputer).

| Candidate observed in MCP | Useful experiment | Current relevant constraints; re-check live |
|---|---|---|
| `seedance_2_0_mini` | Draft motion with approved reference images | 4–15 seconds; 480p/720p; native audio switch |
| `seedance_2_5` | Reference-rich continuity or controlled edit/extension | 4–30 seconds; 480p/720p/1080p; mode-specific fields |
| `kling3_0` | Short acting/camera comparison | 3–15 seconds; std/pro/4k; `sound` uses on/off strings |
| `cinematic_studio_3_0` | Deliberate camera and high-value compositions | 4–15 seconds; resolution and audio parameters |
| `minimax_h3` | Reference/keyframe comparison | 4–15 seconds; inspected resolution option 2K |

These are **benchmark candidates**, not a ranking or evidence that any model preserves Lumi/Tiko. The same reference may work differently across models. Elements support must be checked on the actual tool/model surface; do not inject an Element ID into a media role that does not support it. MCP model IDs are not necessarily Cloud API endpoint paths.

## Best skills by business use

| Skill / technique | CartoonOS use | Other business use | Decision rule |
|---|---|---|---|
| Reference/brand asset creation | Model sheets, props, environment boards | Brand kits and reusable campaign assets | Approved original references and rights first |
| Shot planning + keyframes | One readable learning action per shot | Product demonstration or explainer | Lock composition before animation |
| Voice direction / audio tools | es-419 audition, pronunciation and emotional pacing | Authorized voiceover | Sample review and rights; no assumed voice ownership |
| Thumbnail generation | Original truthful educational cover | Campaign/video covers | Test the promise against actual content |
| Subtitles | Approved transcript and accessible timing | Localized ads/tutorials | Review names, accents, reading pace and safe margins |
| Faceless video | Optional narrator-led explainer experiments | Adult educational marketing | Not the default for character-led Lumi/Tiko dialogue |
| Product photoshoot | Later original merchandise assets | E-commerce product stills | Product detail fidelity and asset rights |
| UGC product/tutorial/website | Separate adult business campaigns | Product demos, SaaS onboarding | Correct skill for actual input; no fake testimonial claims |
| Ad Multiplier | Later, clearly separated campaigns | Independent edits of an owned finished ad | Measure distinct variants; avoid repetitive channel filler |
| Video analysis | Shot breakdown and edit QA on owned outputs | Campaign diagnosis | A model critique supports human review, never proves safety |
| Scene Builder / 3D Jutsu | Later reusable environments/props | 3D product scenes | Adopt only if scene consistency saves measured rework |

Several of these skills are already available in the connected plugin. Do not install every connector. Start with production/reference work, audio, thumbnails, subtitles and read-only analytics. Add an adult marketing workflow only when its business needs it. Your own reusable skill sources live under `skills/`; they are versioned project artifacts, not silently installed into Supercomputer or global Codex configuration.

## Production sequence

1. Load the approved brief, evidence, exact character refs, style traits, ShotSpec and budget.
2. Read live model details. Validate duration, aspect ratio, parameter types and media roles. Save the schema and hash; estimates use exactly the intended payload.
3. Start with one approved still/keyframe and one primary action. Use the cheapest profile that can answer the experiment; do not confuse low resolution with low cost without a quote.
4. Capture quote, cap and reservation before paid submit. If a billable estimate is missing or the current account cannot fund it, stop dispatch and return the missing prerequisite.
5. Submit once. Save provider request ID and correlation ID immediately. On ambiguous timeout, mark `unknown_submission` and reconcile. **The current public API does not accept submission idempotency keys.** A local unique key prevents duplicate application intents but cannot guarantee exactly-once provider delivery. [Errors and retries](https://docs.higgsfield.ai/docs/concepts/errors).
6. Poll by saved job ID with bounded backoff. Handle queued/in_progress/completed/failed/nsfw/canceled; record a moderation rejection without retrying it as a transient failure. Validate all credential-bearing polling URLs against the configured provider origin.
7. Copy outputs into owned private storage immediately, inspect media metadata and SHA-256, record actual cost and QA. Provider outputs are retained for at least seven days and may later disappear. [Lifecycle](https://docs.higgsfield.ai/docs/concepts/requests), [billing and retention](https://docs.higgsfield.ai/docs/concepts/billing-and-retention).
8. Reject defects by reason; change one variable for the next intentional attempt. Upscale/edit/audio treatments are new costed jobs. Never relabel an upscaled raster as native 4K or true vector.

Batch only independent approved candidates, capped by current tool and account constraints (the connected batch tools describe 1–6 requests). Preserve shot ID → request index → job ID. Independent jobs do not remove a dependency on an approved prior end frame. Small controlled batches reduce budget and debugging surprises.

## Reusable shot prompt

Production template ID: `HF-CARTOON-SHOT-v1`. Fill placeholders from the approved brief; do not submit unresolved placeholders.

```text
Shot: {episode_id}/{scene_id}/{shot_id}, {duration_seconds}s, {aspect_ratio}.
Use the attached approved references for {exact_character_versions}.
Identity invariants: {silhouette, colors, anatomy, clothing, signature features}.
Setting and continuity: {approved environment, props, light, previous end-frame facts}.
One primary action: {observable action, with clear starting and ending states}.
Emotion: {want, expression, change visible through acting}.
Camera: {shot size, one explicit camera motion, focus target}.
Learning action: {accurate visual mechanism from approved evidence}, readable focal area.
Timing: {action beats with seconds}; leave room for {reviewed dialogue or silence}.
Audio: {approved es-419 dialogue/voice plan or silent draft}.
Constraints: preserve anatomy and references; no added characters, text, logos,
unreviewed factual details or imitation of another franchise's style.
```

Planning example for S01E01, requiring actual ref attachments before any generation:

```text
5-second, 16:9 static medium two-shot of CHAR-LUMI-v1 and CHAR-TIKO-v1
using their approved references and an approved calm canal observation setting.
Lumi pauses, looks carefully at Tiko, then points toward the observation notebook;
Tiko responds with a curious expression. Keep Tiko's approved gill anatomy intact.
One gesture only, clear faces, uncluttered background, no on-screen labels or new animals.
Silent motion draft. No factual narration or regeneration demonstration in this shot.
```

Narration/dialogue stays in reviewed es-419. Camera terminology can remain English in provider prompts if benchmarks show better adherence. Never use copyrighted character names or style imitation as prompt shortcuts; describe your own formal traits from the mascot bible.

## Prompt repair loop

Record observation → hypothesis → single controlled edit → result → learning. Identity drift: inspect references and reduce camera/pose complexity before adding adjectives. Unreadable learning: simplify background and focal action. Bad hand/prop interaction: separate reach, contact and reaction. Poor dialogue timing: audition/re-record or split shot; never accelerate speech beyond comprehension just to fit a duration. Factual defect: return to evidence and storyboard, not a cosmetic prompt patch.

## Benchmark starter suite

Begin with six representative shots: Lumi close-up; Tiko close-up/gills; duo interaction; walk/point; prop interaction; calm educational visual. Compare two eligible models with the same refs, duration class and rubric, initially one attempt each within an approved small cap. Record results as exploratory; this is too small to establish universal superiority. Expand repeats and the ~20-shot suite from the original integration strategy after observing variance.

Rubric: identity, anatomy/signature, action clarity, camera adherence, continuity, learning correctness, speech intelligibility, safety/originality and technical export. Safety, originality and factual errors fail regardless of aesthetic average. Record reviewer, rubric version, frame/timecode, severity and rejected reason. Compare first-pass acceptance, attempts per accepted shot, accepted seconds, cost/accepted second and median/p95 latency with sample counts. Marketing claims never populate QA scores.

## Supercomputer handoff

Supercomputer is the creative workspace; CartoonOS stores the durable plan and provenance. Open one project per business/channel boundary. Attach the current brief, approved refs and relevant skill only. Each handoff returns shot ID, prompt/version, model/capability, params, quote, job ID, output references, cost, QA and next action.

The remote shell is ephemeral (tool documentation describes roughly ten seconds after a call completes). Feature-detect tools; obtain the upload URL before producing a file; perform download → probe → process → validate → upload in one bounded job; confirm only after successful upload. Never place the only original, credential, canonical script or metric file in that sandbox. Longer background jobs need bounded polling and exported artifacts within the documented lease.

No verified native Supercomputer workflow-event export to this CRM has been established. For now use explicit handoff records; the target adapter writes equivalent events to CartoonOS. Do not promise a live view of every private Supercomputer chat or internal agent trace without a supported export/API and permission.

## Useful requests to start work

```text
Use skills/cartoonos-production/SKILL.md. Prepare S01E01's first Golden Shot
from approved Lumi/Tiko references. Return the exact candidate payload, quote,
missing prerequisites and QA rubric. If refs, account credit or approved cap
are missing, return a blocked handoff; do not invent bindings or results.
```

```text
Use skills/cartoonos-weekly-review/SKILL.md. Review this week's real costs,
workflow exceptions and eligible snapshots. Identify one production bottleneck,
one evidence-backed experiment and the next three unblocked tasks.
```