name: AI-video-iteration-doctor description: AI视频迭代诊断工具：根据生成结果定位动作、运镜、角色、空间、材质、节奏或提示词问题，并给出下一轮最小且明确的修改方案。用于结果不对、需要重试或不知道该改哪里时。 veRSIon: 1.0.0

AI Video Iteration Doctor

Bottom-Layer Reasoning

APPly the think-one-step-further mechanism as a Lightweight check:

• Confirm this Skill is solving the user's real intent, not only the surface request.

• Make the ouTPUt immediately usable and name any unavoidable next ACTion.

• Extract the reusable structure and, when aligned, propagate it to adjacent Prompts, skills, checklists, or workflows.

• Add one guardrail for the most likely next failure while preserving this skill's primary workflow.

Core Intent

Use this skill after a generated image/video result fails. Diagnose the likely cause, choose the smallest next change, and write a prompt patch. The goal is controlled iteration, not rewriting everything every time.

If the user provides an output video, use $ai-video-output-review first to extract timestamped evidence and inspect what is visible. For a still image, inspect it directly. If they only describe the failure, diagnose from symptoms and ask only for missing details that block a useful fix.

Workflow

1. Record expected / actual / first visible deviation / preserved successes in plain language. Do not begin from a global verdict such as ‘not cinematic’.

2. Reconstruct the actual attempt before blaming the model: inspect the generated result, prompt, reference roles, mode, duration, source images, and contradictory constraints when available.

3. Choose one suspected causal layer: input/brief, reference responsibility, mode/parameter, identity/geometry, action/physics, camera, look/material, edit/post, or verified model limitation.

4. State one primary change, what must remain unchanged, and the predicted visible evidence. Add at most two SECondary fixes only when the result is otherwise unusable, and declare that such a round cannot prove causality.

5. If the Same failure family repeats twice or more, pause normal patching and escalate to evidence gathering before writing the next full prompt:

• use $ai-video-output-review for visible timestamped proof when an output is available;

• use $reference-hunting-board / $creative-research-first for missing Aesthetic, camera, action, effects, or sound references;

• for action stiffness, unclear 套招/喂招, repeated movement, or bad body method, use $action-choreography-reference with real movement/action references before redesigning.

6. Write a retry patch that preserves what worked. When a parameter or platform mode is under test, record syntax accepted separately from effect observed; no error is not proof that the intended control was effective.

7. Before writing any negative or context-change wording, run a contamination check. Do not put the failed image, old context, old prop, old location, old style, or old action mode into the negative patch unless it is a broad recurring failure class. Convert the retry into current positive visible facts: what must now be visible, what changes state, what stays locked, and what kind of generic drift/artifact is forbidden.

8. Tell the user what not to change this round.

9. If the user asks for a complete prompt, or the next attempt is likely to be copied directly into Jimeng/Seedance, route the patch through $jimeng-sd2-prompting and deliver one Standalone copyable block with negative constraints inside it. Do not leave the user to merge a patch with an old prompt manually.

10. If a $creative-production-ledger exists, create the retry as a child attempt and record the isolated --change delta; never overwrite the failed parent attempt.

11. After the user confirms the new result, route reusable evidence to $creative-casebook: record the visible result, isolated prompt delta, applicability, side effects, and whether a shared skill should inherit the lesson.

12. If needed, route to specialist skills:

• $ai-video-output-review for timestamped extraction, contact sheets, prompt/reference Comparison, and post-generation acceptance review.

• $cross-shot-consistency-audit when an already selected multi-shot sequence needs an anchor-based drift matrix and repair priority before choosing the retry.

• $bounded-explore-select when many viable candidates need scoring, elimination, feedback, and a stopping decision before any retry is justified.

• $character-continuity-bible for identity/prop drift.

• $cinematic-audiovisual-language for weak camera logic, repeated shots, unclear axis/screen direction, stAGIng drift, or incoherent cuts.

• $screenwriting-story-craft for stiff EMOtional transitions, missing temporal hinge, unclear desire/obstacle, or weak micro-story.

• $action-choreography-reference for weak or unreadable action, weightless weapons, sliding support, missing center-of-mass logic, identical martial styles, impossible braking, or contact without recoil.

• $action-rhythm-editing for pacing and impact.

• $visual-style-aesthetic-direction for cheap/generic/incoherent aesthetic direction or style drift.

• $ai-material-Realism for fake texture, weak render language, missing AO/contact shadows/PBR/anisotropic highlights, or plastic CG.

• $visual-reference-vocabulary for missing vocabulary after the shot grammar problem is understood.

• $ai-video-prompt-preflight when the next prompt is long, multi-reference, action-heavy, or likely to fail again.

• $jimeng-sd2-prompting for final prompt formatting.

• $creative-casebook when the user says a result worked, failed in a repeatable way, should be remembered, or should become a template/rule.

Read references/failure-diagnosis-map.md for symptom-to-fix mappings.

Output Shape

诊断：
主要问题：
最可能原因：

下一轮只改：
保留：
不要改：

提示词补丁：
...

负面限制补丁：
...

结果确认后沉淀：
...

Guardrails

• Change one main variable per retry unless the result is unusable; write the predicted visible evidence before running.

• Do not promote a universal rule from one random A/B result. Repeat a minimal sample or verify on a new scene before cross-project reuse.

• Preserve successful parts explicitly.

• Distinguish observed failure from inferred cause. Say "most likely" when the execution evidence is incomplete.

• Do not call something a model limitation until prompt overload, reference conflict, unsupported mode, asset weakness, and workflow errors have been checked.

• Do not recommend longer prompts by default. Often the fix is removing conflicting instructions.

• Do not fix weak camera, stiff emotion, or position drift by adding style adjectives. First repair shot function, temporal hinge, blocking, axis, or reference roles.

• Do not fix weightless action by adding more shake, sparks, debris, speed lines, or "powerful" adjectives. First repair actor/load relation, support, mass distribution, contact response, and inertia resolution.

• Do not fix fake-looking ACT/POV fights by adding more camera aggression alone. First repair the underlying action relationship, but do not expose the old five-part "attack intent -> preload -> attack-line proof -> forced response -> new state" formula as a visible prompt template. Use it only as an internal diagnostic check. For SD2/Seedance copyable prompts, translate it into the tested natural cause-effect wording: "when A does X, B is visibly forced into Y; A uses that result to do Z; the spacing/angle/elevation becomes N." The viewer should see contact, receiver response, and changed state, not a tactical essay.

• Do not fix weak aesthetics with name-salad references. First choose medium domain, style family, color ownership, rendering hierarchy, and anti-cheapness constraints.

• Do not fix material/render failures by dumping keywords. Attach AO, PBR, anisotropy, GI, SSS, or volumetrics to visible surfaces and light conditions.

• Do not keep writing full new prompts after the same failure repeats. First inspect the generated output, compare it with at least one suitable reference or known-success case, and DeFine the smallest evidence-backed change.

• If a 10-15 second CLIp fails, suggest splitting into shorter clips.

• If identity or prop drift is the issue, put locks earlier and reduce camera/action complexity.