Requesting generation feedback
Overview
Pruna video and replace jobs are expensive. This discipline skill governs when to pause, what to show, and what wording to use before any paid generation or final audio mux — regardless of which workflow skill is active.
Canonical phase rules: staged-generation-gate.md
Per-workflow runners and flags: workflow-feedback-gates.md
When to Use
- Before the first
POST /v1/predictionsin a session - Before Phase B video (
p-video,p-video-avatar,p-video-replace,p-video-animate) - Before Phase D assembly with background bed or full-song mux
- When the user says "just run it" but stills or clips have not been reviewed
- When art direction is ambiguous (visual mode, cast, continuity, bed yes/no)
When NOT to use as an excuse to stall: the user already replied approve plan, approve stills, or approve clips for the current phase — proceed with that phase only.
Red flags
Stop and ask (or show assets) if any of these are true:
| Red flag | Risk | Required action |
|---|---|---|
| Plan not presented or no approve plan | Wrong story, cast, or style bible | Phase 0 — scene table + sample prompts |
| Stills not shown since last prompt edit | Silent no-op regen; wasted video credits | Phase A — paths in stills/; wait for approve stills |
| TTS / song not listened when narration drives video | Bad pacing, wrong lines in lip-sync | Phase A2 — audio/narration_*.mp3 or song.mp3 |
| Same turn: plan approval + video | User never saw plates | Split turns; never batch |
| approve clips missing before concat + bed | Bad VO buried under music | Phase C/D only after clip review |
| Visual mode, cast gender/voice, or continuity unclear | Identity drift, wrong pipeline | Ask; do not guess |
Using --yes-skip-*-gate without user asking for automation |
Bypasses human review | Confirm explicitly |
| Regen prompts without deleting stills/clips | Old assets reused | Delete targets or --fresh / --regen-* per staged-generation-gate.md partial regen table |
voice_script revised but avatar sources not deleted |
Lip sync / dialogue mismatch | Delete sources/ + clips/ for that scene |
POST /v1/predictions without random seed ritual (SSoT) |
Duplicate outputs; copied example strings | Generate and state a ritual string first; log ritual_seed |
PRUNA_API_KEY or REPLICATE_API_TOKEN missing |
Cannot run API or runners | Stop; send signup links from api-credentials.md — Pruna dashboard and/or Replicate API tokens |
Workflow
- Classify the next phase — plan (0), stills (A), audio prep (A2), video (B), assembly (C), final audio (D). See staged-generation-gate.md.
- Present the right artifact — table for Phase 0; file paths or thumbnails for stills; playable audio for A2;
clips/for B. - Use the wording templates from staged-generation-gate (approve plan / stills / audio / clips).
- Cost warning before Phase B — state that
p-video/ avatar / replace / animate are paid. - Run only the approved phase — default runners use
--phase stillsor--phase song; unlock video with--approve-stills(orgeneration_status.json). - Partial regen — fix one asset, not the whole pipeline; follow the decision tree in staged-generation-gate.
Runner flag cheat sheet: workflow-feedback-gates.md (Runner flag reference + per-workflow commands).
Common mistakes
| Mistake | Fix |
|---|---|
End-to-end --phase all on first run |
Default phased flow; skip gates only when user requests automation |
| Showing manifest JSON instead of media paths | User reviews JPEG/PNG/MP3/MP4 |
| Approving "looks good" without listing paths | Name stills/hero.png, scene ids, or clip filenames |
| Mixing bed before clip review | Concat first; bed after approve clips |
| Assuming regen picked up prompt edits | Delete affected files or use --regen-stills / --regen-clips |
Related
- staged-generation-gate.md — phases, templates, partial regen
- workflow-feedback-gates.md — workflow index + CLI flags
- generation_gate.py —
generation_status.jsonhelpers - Scenario hub: pruna-generative-pipeline