community-webtoon-producer

Community Webtoon Producer

Objective

Produce a readable Korean community webtoon without hallucinating the community, skipping causal story work, or confusing successful image generation with user approval.

Primary Quality Rule

Treat shared context and understanding alignment as a production input, not as conversational polish. A precise micro-prompt cannot recover a missing whole-story reason, reader-emotion sequence, recognizable motif, fact/MSG boundary, or user-observed mismatch.

Preserve context that may look irrelevant to the immediate command when it changes creative judgment. Re-state the enough context to reconstruct the intended judgment inside every downstream packet that asks another model or pass to decide. Paths, hashes, references, and previous images prove state or appearance; they do not explain why the work should look or feel that way.

This rule comes from repeated quality improvement after failed prompt-only passes. It is not an abstract collaboration philosophy.

First Read

From the repository root, read:

1. AGENTS.md
2. docs/BEFORE_YOU_START.md
3. docs/QUALITY_CONTRACT.md
4. the active project's PROJECT.md

For creative collaboration, read docs/CREATIVE_ROLES.md and docs/CREATIVE_CONTINUITY.md; carry the relevant expertise, WHY and current delegation into the actual model request. Initialize a project with harness/scripts/init_project.py when the user is ready to develop the agreed material into production.

Required Workflow

1. Collect the primary post and a narrow sample of surrounding reaction sources. Record URLs and observed times.
2. Separate primary_fact, mood_source, MSG, and do_not_use in the source ledger.
3. Discuss and obtain approval for the sweet spot: Director intent, community reader expectation, and comic pacing.
4. Write and preserve a rough Director/producer beat skeleton. When a deliberate blank is part of the brief, leave it open instead of pre-solving it before the creative pass.
5. Give the skeleton and the full source/fact/MSG boundary to Gemini as a blank-filling or staging pass. Preserve the raw Gemini output beside the raw skeleton; do not overwrite either artifact.
6. Review the two artifacts with the Director and record each substantial proposal as adopt, adapt, or reject. Codex may enter the work to correct facts or continuity; that correction can remain visible when it serves the comic.
7. Consolidate the accepted result into one canonical conte. Give every cut handoff_from_previous, reader_first_sees, character_realizes, and push_to_next, then reduce density only after causality is visible.
8. Require a human to read, edit if needed, and explicitly approve that canonical conte. Record the raw skeleton, raw Gemini pass, joint review, and human-approved canonical version in editorial_review_lock.csv.
9. Promote the human-approved conte into page/cut packets with stable packet IDs.
10. Classify every page as independent_page, same_scene_continuation, or reused_shot_variation. Page order and dialogue progression alone do not prove visual continuity.
11. Lock visible text source class and approval state.
12. Lock visible text routing: owner, role, attachment, read order.
13. Lock visible cast, reference roles, hashes, and co-appearance rules. If a recurring character has no human-approved identity reference, stop and request one instead of claiming identity consistency.
14. Write one self-contained provider-bound prompt per page using the generation contract.
15. Run validate_project.py --stage pre-generation --strict.
16. Wait for explicit user authorization to generate.
17. Use ima2-gen as the default runtime and record the actual runtime, inputs, and outputs. After authorization, generate at least three independent candidates; prefer four to six for acting, expression, action, reveal, or atmosphere. Use a single candidate only for a technical probe, explicit cost limit, or direct user request.

• If one independent job fails while later work remains, carry only the missing job into the next live wave immediately, ahead of ordinary jobs, without rebooting or cancelling successful in-flight work. After the final wave, make one isolated tail retry and leave any remaining failure explicit.

1. Hand every candidate to the user for visual judgment. Report only objective facts unless critique is explicitly requested.
2. If the human-selected page already has the right structure, acting, cast, and story beat but has repairable lettering or a few local visual defects, optionally use harness/templates/post_selection_page_finish_request.md. Preserve the selected original and hash, replace the complete final visible-text inventory from the source/routing lock, describe only cut-local visual repairs and the keep list, and save the edit as a new candidate for human comparison. Regenerate instead when structure, causality, identity, fact/MSG boundary, or core acting is wrong. This path remains experimental until density trials prove it reduces time to usable result.
3. Move unselected or superseded candidates out of current. Do not delete evidence.
4. After final approval, freeze final composites, approved sources, and actually used materials with SHA-256 manifests.