ayase0307/ai-comic-drama

AI Comic Drama Production (Pure Agent Version)

Startup Rules (Always Follow, No Exceptions)

1. First, run python scripts/h3d.py next --root <project>. Follow its single printed instruction and read the specified reference. Do not arrange your own workflow or skip steps.
2. User decisions must be recorded with h3d decide <id> <value> --quote "<user's exact words>" into decisions.json (art style, aspect ratio, dialogue policy, design engine, acceptance policy, editing method...). Do not only write in notes, and do not decide for the user.
3. Do not manually edit review/project-state.json. Its state can only be written by build_stage_review.py; manual edits will be flagged as blocking by h3d next. If Stage 6 technical QA is PARTIAL and the user has seen and approved it, use --approve-stage 6 --override --reason "<user's reason>".
4. Stage 6 per-segment conclusions must use h3d verdict to write into review/verdicts.json (pass/warn/fail + one-sentence reason + evidence image), classified per acceptance_policy. 3a. The image generation backend is chosen by the user, not assumed by the agent: Before generating images in Stage 2 (design) or Stage 4 (composition), run h3d imagegen --detect --root <project> to list detected available tools (with reasons for unavailable ones) and ask the user to choose. Ask even if only one tool is detected. After the user responds, run h3d imagegen --set <backend> [--stage 2|4] --quote "<user's exact words>". If h3d next encounters a stage where no backend is selected, or a selected backend is now unavailable, it will change to ask you to inquire (see references/tooling.md §2.4.1). 4a. Decisions that need user input are generated via h3d inbox into a decision inbox for the user (only exceptions are listed). For stages the user has authorized, follow h3d next using h3d approve --auto. Stage 0, character design images, and the voice of each character always require human decision (see references/tooling.md §2.6).
5. Do not run two video runners simultaneously. When the ledger has waiting/running jobs, only collect results (h3d collect), do not submit new jobs.
6. When encountering PARTIAL or duration_drift, stop immediately. Follow references/tooling.md §4 for disposition and report to the user. Do not bypass on your own.
7. Default final product format = long-form comic drama: Telling the story clearly is the top priority. Every key event from the original work must cover cause → process → consequence (including motivation, time and place transitions). Do not compress events based on per-episode duration. Default is no episode split; a single continuous long film. Short-form drama / vertical-screen viral short episodes (1–2 minutes, strong hooks) are only used when explicitly requested by the user (h3d decide series_format short). Any events that are compressed, skipped, or not fully explained should be re-added by the Agent based on the original work (event_coverage.py). Do not cut them yourself or ask the user.
8. Do not ask the user about script adaptation: Always follow references/adapt/dialogue-first-adaptation.md (preserve all original dialogue; only narrations that are rephrased dialogues can be rewritten and marked invented; narration = protagonist's internal voice VO ≤30%; no cap on episode length; episode breaks should stop at natural story points and record the original anchor + reason; pass all checks in adapt_gate.py before delivery).
9. Before starting work, check the skill version: If h3d --version differs from the creation version recorded in the project's review/project-state.json, synchronize the skill first before starting. The same prompt may have different thresholds across versions (single sentence 15→18 chars, speech ratio 60%→90%, minimum shot length 2.0→1.2 seconds, added D2 density and D4 openings). Rewriting with old rules will result in wasted effort. The skill copy in the installation directory may not be a git repo. If the source repo is local, pull it first, then replace the entire skill directory, backing it up first.
10. Fixes should be batched, not sentence-by-sentence sign-offs: Fix all segments with the same type of issue (same sentence structure, same package problem) at once, then rebuild → h3d bind → h3d compile → re-sign-off Stage 4/5 (batch = first check each shot against its shot_intent to confirm the same root cause, not bulk fix by warning type). Re-signing after every single edit will consume the entire review cycle, and Stage 5 will block if layer2 and segment fingerprints are out of sync. Rebuild clears sketch_* fields; snapshot by shot_id before rebuild, write back after. If the number of segments or segment IDs change, sync the REMIX storyboard (PUT /projects/{id}/storyboard): The interface/review room's "Expected N shots" and chapter grouping read from this document, not from segments-*.json. If not updated, it will always show the old plan (e.g., 133 shots vs actual 18 segments). Before swapping, use GET /jobs?project=<id> to check each shot ID for no active jobs. If none are active, replace the entire file directly. When the user says "draw cards" or "arrange the review/revision/cycle yourself" = run a full autonomous loop (submit → collect → draw frames for review → revise → draw again), only stopping at human checkpoints (Stage 6 approval, user-written override reasons). Report with the unit of "round + 4-grid evidence images + what was fixed", not reporting after each single drawn segment.
11. Reference package overrides prompts: Whatever the REMIX package's master shot contains, will appear in every shot (e.g., glowing buildings, backpacks, night sky). To change, swap the package or regenerate the master shot. Do not use negative sentences in prompts to force removal — mentioning a prop means it appears. Details in references/remix-draft-loop.md.
12. Approved and final versions are stored separately; changing versions requires re-verification: videos/<segment>.mp4 always holds the version that has passed QA and been signed off. pika768 finals, upscales, and re-drawn candidates go into subdirectories like videos/final-768/. To promote a final to approved version = re-run h3d qa / motion / --continuity → rebuild Stage 6 snapshot → user signs off again. Do not just overwrite the file to change versions; this will break the signed-off snapshot-evidence link.

Overview

Depending on the delivery scope specified by the user, this skill completes novel adaptation, script, director's storyboard, prompts, or generates media using available image/video tools. Textual planning can be delivered independently; media generation requires corresponding tools or a connected backend (FreeVideo is an optional adapter, not implying every Agent has built-in video generation). If no generation tool has been actually called, do not claim that images or videos have been produced. Novels must not be used directly as scripts; always go through Stage 1 adaptation first. When inheriting an existing project (not starting from scratch), audit before proceeding: First read references/audit-inherited-project.md. Do not continue writing along the previous agent's storyboard — a broken upstream will cause every shot to be redone. Before determining "duplicate assets" or "content errors", always cross-reference the bible and original line numbers, not just the images or filenames.

This page is the Operations Homepage: what to do in each stage, what not to do, which instruction to run. Principles, templates, practical decision records, and the full tool list are in references/skill-guide.md (the full original SKILL.md, chapters unchanged), which h3d next will point you to the relevant part.

When to Use

• "Turn this novel/story into a comic drama, dynamic comic, or short drama" / "Adapt the novel into a comic drama script"
• "Write storyboards/shot list/video prompts" / "Design character settings/character bible" / "Turn product material into a product short video"

Per-Stage Operations Homepage

Stage	What to Do	What Not to Do	Which Command to Run	Read First
0 Toning	Ask all toning questions at once (genre, medium, art style, aspect ratio), each with recommended values; Dialogue policy is fixed: Original dialogue first, no need to ask; Default final product format: long-form comic drama (don't ask about episodes/duration; only use h3d decide series_format short if the user actively requests short drama); Default art style recommendation: No. 14 "Ancient-style realistic thick paint, cinematic feel" (with sample images, can be changed)	Don't decide for the user; don't skip volume estimation; don't default to short episodes	init_project.py, h3d decide style/aspect_ratio	references/pipeline.md, references/styles.md
1 Adaptation	Retain grading; ensure event coverage completeness (each event: cause → process → consequence; missing parts re-added by Agent); dialogue-first adaptation and episode splitting based on dialogue budget; paragraph structure; causal debts; pass all adapt_gate.py checks before delivery	Novels used directly as scripts; events compressed for duration; voluntarily cutting events	event_coverage.py, dialogue_budget.py, adapt_gate.py, h3d next	references/adapt/dialogue-first-adaptation.md, references/adapt/novel-to-manju.md, references/adapt/structure-and-pacing.md §Zero
2 Design Images	Detect image generation tools, then let the user choose (ask even if only one is available); generate main character images first, re-generate after self-check; lock the tool for the entire episode after decision; Every script effect must have its own B- effect card and reviewed reference image (no image, no entry to Stage 5, E56)	Assume the image gen tool by yourself; change tools without side-by-side comparison; auto-approve (faces and costumes always need human review)	h3d imagegen --detect / --set (writes design_engine), review page	references/consistency.md, references/styles.md
3 Script	Scene-based adaptation; dialogue budget (character speech rate 4.0 / 4.5 / VO 5.5 chars/sec; speech ratio 75–90%); adaptation differences; Blind review before finalization (fresh subagent given only script materials: build_blind_review.py --stage 3, rules in references/blind-review.md)	Submit over-budget scripts	estimate_duration.py --h3, sync_script.py	references/adapt/screenplay-format.md
4 Storyboard	For each shot, write shot_intent (what the audience needs to understand) first, then technical values, with a reason for each choice in choice_why; Camera movement based on beat emotion, consult references/camera-moves-guide.md to select move_preset (reason written in storyboard, compiler expands, move_lint warns); Every segment: character ledger, dialogue shots include eye line, segment seams include final state and camera position change (continuity-ledger.md); Ref2VA defaults to one clearly identifiable face per shot (single-person crop, not restricted when using character card remix), no people in wide shots; focal length and distance written as pairs	Enter compilation without corresponding composition PNG; bulk fix by warning type; modify approved content to pass checks (e.g., crowd replaced with empty seats, filling empty background); composition images change people or miss key props	Missing composition: h3d imagegen --detect / --set --stage 4 (user chooses Codex storyboard / Qwen pencil 9-grid / hand-drawn, writes storyboard_engine) → h3d next gives tool-specific instructions (Codex needs --model <user-specified model>), validate_storyboard.py	references/storyboard/shot-intent.md, references/storyboard/cinematic-director.md, references/storyboard/continuity-ledger.md; for combat/skills/effects also load h3-combat-impact (see next section)
5 Prompts	Manually write Layer 2 five_part_en per shot (Camera / Motion / Performance / Light / Sound; Motion and Performance must be detailed; each clause must correspond to a visible or audible result; do not rewrite appearances already in reference images; write positioning, eye line, start/end posture, key props); write after reading storyboard, composition, and previous shot state; the compiler only assembles the six-part format; Before submission, run prompt_sketch_audit.py to check each clause against the composition image (for fl2v segments, first run frame_pair_qa.py to overlay images)	Write code to generate the five subsections; overwrite segments after compilation; use metalinguistic terms (authored/approved) or prohibited phrase stacking to get by; ignore warnings without acknowledgement	h3d bind (first generate reference images and bind), h3d author, h3d compile --check, h3d catalog --check	references/prompt-authoring.md, references/h3/prompt-formats.md, references/tooling.md; for combat/skill segments also follow h3-combat-impact
6 Video	Submit → Collect → QA → Motion QA (optical flow freeze/stutter → automatic speed change candidate list) → Inter-seam continuity QA (actual cut points + seam composites + one-side freeze candidates, warnings only) → --continuity-pairs answer the 6-item continuity checklist per pair (location / prop / screen side / costume / light / effect) → Per-segment conclusion → Inbox	Run two runners at once; touch other people's jobs; send drifting footage to QA; use planned duration to pick frames for seam checking (first check --continuity to find actual cut points)	h3d probe, h3d run --plan, h3d qa, h3d motion, h3d qa --continuity, h3d verdict, h3d inbox	references/tooling.md §2.5–2.8 (seam QA see §2.8.1), references/remix-draft-loop.md (remix card draw → review → revision), references/speed-ramp.md, references/skill-guide.md "Stage 6 Generation"
7 Subtitling & Editing	Post-production speed changes on approved action segments (original video unchanged, dialogue at original speed); loudness; finalize each character's voice once; single editing project for the entire episode; Blind review before delivery (build_blind_review.py --stage 7 frame collage, rules in references/blind-review.md)	All dialogue, narration, and voice-over generated natively by H3 (no external TTS); set voice for the user; overwrite original video; speed change in dialogue areas	h3d ramp, h3d concat plan/build, h3d assemble; editing lens (suggestions only) scripts/craft/edit_review.py	references/tooling.md §2.7, references/skill-guide.md "Stage 7", references/edit/manju-editing-playbook.md (cut points / transitions / effects / subtitles / title cards / music ducking / loudness), references/craft/murch-cut, references/craft/hamilton-action

Official stage approval is based on review/project-state.json, and can only be done via build_stage_review.py --approve-stage (with user's exact words) or h3d approve --auto (for stages the user has authorized). Video generation pipeline: Default Ref2VA (FreeVideo, single-person crop). When the user switches to the console REMIX character package, finalize with h3d decide video_backend remix, change h3d bind to the character package, h3d compile with remix binding to produce the REMIX three-column format (style anchor at start, <Picture N> / <Video N> numbered per package), audit branches by backend (the submission adapter is not yet connected; FreeVideo's submission side will reject remix segments; see references/tooling.md §2.3.1c). Stage 5 review page automatically runs refcheck/staging for ref2va segments and lists them as reminders (remix segments run within audit). For manual run, use h3d refcheck (characters/creatures/props/crowds mentioned in prompts must be in the package; remix blocks, ref2va only hints) and h3d staging (positioning left to right, eye line, not facing camera; default warning, only blocks with staging_gate strict). See references/tooling.md §2.3.1d.

Stage 5 segment-level approval: Stage 5 approval records a fingerprint for each segment. If only some segments are changed after approval → only the changed segments (including added/deleted segments) need re-review; unchanged segments remain approved and can still be submitted (h3d submit / h3d run will automatically skip or reject segments needing re-review). Changed segments always require manual re-review: build_stage_review.py --project-dir . --stage 5 --segments X,Y generates a re-review page → after user passes, h3d approve --stage 5 --segments X,Y --user-note "<user's exact words>", do not use --auto. Segments that already have videos and are changed appear as stale (prompt_stale) in h3d deps --status. If old snapshots or shared parts (segment file shell, unbind segment ack, upstream stages) change, full re-review is still required. Details in references/tooling.md §2.6. Text-only drafts can be delivered within a specified scope without proceeding to media generation: use h3d decide delivery_scope script-only|storyboard-only|prompts-only|full --quote "<user's exact words>" to finalize the delivery endpoint (no decision = full pipeline). h3d next will report done when reaching the endpoint. With no_image, Stage 2 only requires text-based design cards. Note: no_image only lifts Stage 2; Stage 4 onwards still require composition images. The endpoint must be before Stage 3 or as decided separately. The deliverables and HTML review page mapping for all 8 stages are in references/skill-guide.md "8-Stage Pipeline and HTML Breakpoints". Detailed templates in references/pipeline.md. HTML review pages for each stage are now debug pages (for binding review snapshots). The item handed to the user is the h3d inbox inbox; specifications in references/review-page-spec.md.

Combat / Skills / Effects Segments: Also Load h3-combat-impact

When the storyboard or prompts involve combat, fighting, skills, special moves/finishing moves, one-shot kills, transformations, magical explosions, special effects, impact frames, screen shake (or any L3+ hits), simultaneously load h3-combat-impact (ayase0307/h3-combat-impact), using both:

• Stage 4: Follow its mandatory workflow to design character skills, fight choreography (exchange/consequences), effect layering and camera movement, and list impact points, each with a selected impact frame path (P0–P3). The director-level lint (director_lint.py, warning) will remind: new location needs establishing shot first, dialogue should not cross the 180-degree axis, reaction shots after reversals/bad news, eye-line matching must specify what is being looked at, inter-segment seams use action editing or occlusion cuts (reference 24frames.art cinematography language, adapted).
• Stage 5: Write the five_part_en for combat segments according to its decision cards — finishing moves default to "impact frame chain" segment (template F); do not write H3 slow-motion words or frame count controls like near-freeze / hangs poised / like a meteor / for two frames. Write the impact frame plan (timecodes, impact point, force, path, assembly multiplier) into the segment's storyboard notes / delivery notes (segments currently have no dedicated field).
• Stage 7: Approved combat segments use the impact frame plan for speed changes (h3d ramp, or the skill's impact_frame.py ramp) or compositing (impact_frame.py composite). Dialogue areas remain at original speed.
• Shot length: Consecutive impact frame segments and short impact shots can be shorter than h3-video-prompting's "each shot ≥ 2 s"; follow h3-combat-impact.

This repo's supporting materials: references/storyboard/xuanhuan-combat-contract.md (combat branches, action establishment conditions, setting, three impact layers, camera speed terminology, finishing move chain) and references/storyboard/vfx-design-and-review.md (effect design cards, face preservation defense, visual inspection and rework, local correction methods). Both are distilled from external skills, not tested by H3.

This is a referral, does not change this skill's gates: validate_storyboard.py, h3d compile --check, h3d catalog --check must still pass.

Four Deterministic Checks (Part of Delivery)

Duration quanta (17n+5, using metadata.frames for submission), evidence grading (Stage 6 requires VERIFIED), segment-level dependency ledger (h3d deps), failure catalog (h3d catalog --check / h3d ack). duration_drift, PARTIAL, already waiting job all mean "cannot continue now"; disposition is in references/tooling.md §4.

Supervision (Requirements Ledger, Run Before Review and in Parallel with Blind Review)

h3d decide only records formal decisions. User casually expressed preferences not in decisions will be forgotten, and no check asks "has the user's past remark been forgotten?". Supervision adds this layer: python scripts/build_requirements_ledger.py --root <project> builds the requirements ledger (exact words from each decision in decisions.json + inbox pending cards + --quotes manual extracts), --stage N / --all / --inputs optionally builds a check pack containing "only materials, no conclusions", a fresh subagent checks each item against the output (templates and restrictions in references/supervision.md). Same positioning as blind review: process requirement, not a gate; does not modify the h3d state machine. Not passed should be addressed before review. First step when inheriting an existing project: use --all to accumulate checks (in conjunction with references/audit-inherited-project.md).

Resources (Index)

• references/tooling.md — All h3d commands and checks (read before running any command); §2.5 decisions/per-segment conclusions, §2.6 inbox and auto-approve, §2.7 runner/probe/voice-record/edit.
• references/skill-guide.md — Full original SKILL.md: Core Principles 0–14, project structure, Stage 6 hands-on decisions, Stage 7 voice layering, full script list.
• references/pipeline.md — 8-stage detailed process, question lists, templates; three-layer voice decisions and loudness standards.
• references/review-page-spec.md — Review page specs: what is handed to the user is the decision inbox (layout order, required card fields, prohibited items, 8-stage reading view 6 blocks). Old stage pages are for debugging only.
• references/blind-review.md — Blind review (fresh-eyes): material pack with only materials, no conclusions, and process (mandatory for Stages 3/7; scripts/build_blind_review.py).
• references/supervision.md — Supervision (Requirements Ledger): check user statements one by one against output (run before submitting each stage; scripts/build_requirements_ledger.py).
• references/document-standard.md — Stage document standard format: status card + body + revision history, write using h3d doc, do not edit manually; "You decide" items automatically go to inbox.
• references/audit-inherited-project.md — Handling projects from other agents ("you take over" / "the previous agent did a bad job" / "what suggestions do you have"): audit before proceeding, mechanical generation fingerprint threshold, three ways dialogue pipeline can break, REMIX audit, rebuild sequence and pitfalls (re-bind, official camera movement verbs, sketch snapshot restoration, composition images from unsupported backends).
• references/remix-draft-loop.md — Remix card draw cycle (Stage 6): pika480 4/6 step choices, five symptoms of reference package overriding prompts, three things to check in review, three common false alarms (low saturation / border lines / tone jumps), batch revision + re-sign, collection and finalization order.
• docs/skill-roadmap.html — Human-readable dynamic roadmap (8-stage actions, deliverables, who decides, stop signals); when the process changes, synchronize its embedded roadmap-data (tests/test_skill_roadmap.py reconciliation).
• references/adapt/ (adaptation), references/storyboard/ (director's storyboard), references/consistency.md (asset consistency), references/styles.md (style locking anchor), references/freevideo.md, references/model-profiles.md, references/minimax-h3.md and references/h3/ (H3 detailed reference, synced from upstream h3-video-prompting), references/prompt-playbook.md, references/h3/camera-moves-handbook.md (upstream camera moves handbook: 17 styles + Loop Forge 6 styles + xianxia bullet time, Colab-tested annotations), references/craft/ (creative lens, suggestions only, no gate changes), references/edit/ (Stage 7 comic drama editing manual, Concat/CapCut draft investigation).
• Combat/skills/effects/impact frames: external skill h3-combat-impact (ayase0307/h3-combat-impact), see "Combat / Skills / Effects Segments" above.
• references/revision-and-retry.md — Revisions must include change list + linked conclusions; same shot can be re-generated a maximum of 2 times before back-to-layer diagnosis; acceptance comparison principle.
• references/storyboard/continuity-ledger.md — Character ledger, storyboard review checklist, three position types, positioning table, shot-reverse-shot, seam state, reference image usage.
• references/storyboard/spatial-consistency.md — Spatial consistency: scene master shot drawn on floor plan, wall assignment layer (optional scripts/storyboard/wall_map.py), composition board with master shot as base, shots with only creatures tied to in-store reference (Lao Yi Experiment 2026-10, directional evidence); §6 white model division and precision tiers (Foyege, not tested by H3).
• Format: references/schemas/ (decisions, verdicts). Hard rules for agents also in AGENTS.md.