papperrollinggery/Paperrolling-DIRcreative-SKILL / skills/dircreative
dircreative
VerifiedUse after explicit $dircreative for film concepts, stories, scripts, director or creative groups, shot/asset planning, reference-video analysis, and image/video prompts. Real media execution requires the corresponding authorization. Validated ADCO handoffs are supported. DIRcreative source maintenance is outside this Skill.
references/video-distillation.md
View on GitHub# Video Distillation and Effect Comparison
Studio owns the report, not new-film generation. Tools extract;
`film-breakdown-distiller` owns craft. Start from a media receipt and enter
production only when requested.
## Core creative acceptance loop
For every important decision, retain:
1. **Why it works:** source timecode, observable event, intended information,
emotional, spatial or material effect; distinguish fact from interpretation.
2. **Why we need it:** the original film's objective. Discard irrelevant reference
qualities. Comparable quality does not mean the same plot, shots or spectacle.
3. **Current gap:** compare actual script, timed boards, images or sound against
that effect. Name the failure and affected shots, not “less cinematic”.
4. **One correction:** change the dominant cause; preserve successful layers.
5. **Recompare:** inspect the revision using the same criterion. Record evidence,
change rationale and improved / unresolved / untestable outcome.
Reuse the exact `criterion_id` for the same effect across comparison cycles.
Schema PASS is not artistic certification. Verify independent findings. Stop at
scoped acceptance, no gain, or unavailable evidence/authority; do not
cosmetically regenerate adequate assets.
## Evidence intake
- Inspect the named installed downloader/analyzer first. For public Douyin use
`extract-douyin-transcript`; retain resolved ID, sanitized URL, source-to-file
manifest and media validation. No access bypass.
- Prefer the user's Video Evidence Workbench / `video-analysis-mvp`; inspect its
CLI, use a dedicated output directory, explicitly set the authorized measured
duration cap. Do not modify its source/config or silently upload to providers.
- If its provider is unavailable and host-model substitution is requested, read
its actual system prompt, field schema and synthesis code. Inspect all selected
frames and validate companion annotations with its pure validator. Record the
real executor; never forge provider receipts or human approval. Keep neutral
observations separate from cross-shot craft interpretation.
- Verify original master hash/dimensions separately from review-copy dimensions.
Candidate machine cuts may include flashes, dissolves or duration splits.
`scene_no`, `sound_sync`, `music-led`, story beats and tempo can be heuristics;
never treat these populated fields as observations or exact shot counts.
- Inspect the whole timeline with appropriate samples and denser sequences for
motion/cut claims. State exactly what was perceived: stills, frame triplets,
playback, waveform, ASR or descriptions. Waveforms cannot identify instruments,
delivery, stems or lip sync. Without listening, sound interpretation is limited.
- Prefer cached local ASR. Preserve raw output; check captions, names, repetitions,
time reversal, beyond-duration segments and the tail. ASR can hallucinate over
music. Keep correction provenance; never call an unchecked transcript verified.
Use `scripts/dircreative_video_distill.py prepare|validate|render` as the local
bridge; see `--help` and README.
The helper writes an **unknown** draft; Codex authors the analysis.
Without Workbench, shots stay unavailable. Validation covers structure/bindings,
not truth or quality. Agent review cannot impersonate human approval.
## Report and transfer
Deliver a readable report, indexed evidence and structured study. Cover requested
axes; unknown is valid, fabricated certainty is not:
| Axis | Observe → production control |
| --- | --- |
| Narrative | desire/obstacle/causal change → premise, beat map, payoff |
| Dialogue | words, tactic, response, pause → voice, subtext, timed line |
| Directing | gaze/status/action/reaction → intention, blocking, coverage |
| Cinematography | scale, angle, perspective, subject vs camera motion → axis, framing, motivated move |
| Color/light | hierarchy/direction/contrast over time → look bible + color script |
| Texture | material, sharpness, bloom, grain → separate surface/optical/post controls |
| Performance | micro-action, breath, delay → playable performance score |
| Editing | cut reason, duration, ellipsis → timed boards, handles, reactions |
| Sound | audible speech, ambience, foley, cues, silence → cue sheet, stems, mix hierarchy |
| Continuity | identity, space, object state, damage → canonical owners/state transitions |
| VFX | trigger, change, consequence → key states, plate/compositing plan |
| AI workflow | documented methods only → plausible routes, tests, failure rewinds |
Exact camera/lens/stock/filter/model/seed/prompt/budget/team cannot be inferred
from appearance. New-project settings are design choices. Separate soft light,
halation, bloom, depth of field, grain and codec softness; proposals are not facts.
Each mechanism needs evidence IDs, problem, causal mechanism, concrete controls,
use case, misuse boundary and an inspectable check. Preserve effects while
replacing distinctive people, dialogue, events, world rules, composition sequences,
props and music. Do not package the reference's protected full expression.
## Research and controlled self-iteration
Investigate a specific unresolved mechanism. Check existing code/skills, then
current official model docs and first-hand practitioner evidence (ASC, animation
editors, official ComfyUI workflows/source repos). Community claims are leads
until tested. Record URL, access date, claim, source type, applicability and limit.
Stop when evidence is sufficient.
Classify DIR gaps as missing capability / existing but unconnected / weakly
specified / untested. Map each to its source owner. Existing shot/audio/continuity
fields are not missing modules. Prefer a reference, adapter or fixture over a
new orchestrator. Analysis alone produces improvement candidates, not mutations.
When iteration is authorized: baseline -> hypothesis -> minimal branch patch ->
positive/negative tests -> effect recomparison -> independent review -> retain or
revise. Stay inside authorized repo/scope. Runtime never rewrites installed
Skills, global config or unrelated projects. Preserve legacy/model routes.
Commit, installation and publishing need their own authorization. Self-iteration
is a bounded invocation, not a recurring automation or background promise.
## Ordered execution
Finish source inspection and the report before upgrading from its findings.
Finish scoped upgrade tests and independent review before a requested original
production trial. Report an unmet phase honestly; do not pre-generate “final”
assets to simulate progress. Interface research may produce a later PRD, but
must not displace the authorized film/asset deliverable.
## Original film through preproduction
After study/upgrade review, explicitly requested original-film work transitions
to the existing film-development, asset-foundation, shot and model owners.
Complete actual per-shot/scene/identity assets, sound and script, exact timeline,
generation units, reference slots, prompt hashes and attempt ledger. A board
sheet never replaces individual frames; plans never count as generated assets.
Keep planning boards separate from clean inputs and unavailable audio out of slots.
Run the core effect-comparison loop on the actual outputs. Movement, lip sync,
temporal consistency, final edit and final mix remain untested before video.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |