SuperAPI AI Directory
Partner中文AI newsFavorites
Skillswebtoon-orchestrator

webtoon-orchestrator

Design & creation

View repository

Updated 10-16 04:57

About this skill

Webtoon Orchestrator — Webtoon Production Team Coordination

An integrated orchestrator that produces a single webtoon episode from trend research to a finished viewer. It sequentially operates 27 specialized agents as 4 stage-based teams.

Execution mode: Agent team (reconfigured per stage)

Since only 1 team is active per session, for each Phase create a team with TeamCreateand when the work is done clean up with TeamDelete, then create the next Phase team. The previous team's outputs remain in _workspace/and are carried over by the next team via Read. For every team member spawn and Agent call, specify **model: "opus"**.

Agent composition (27 members, 4 teams)

팀 Team member Role Skill Main output
Research team trend-scout Genre/trope trends webtoon-trend-research 01_research/trend-scout.md
platform-ranker Platform rankings·serialization structure webtoon-trend-research 01_research/platform-ranker.md
audience-analyst Readership·churn·immersion webtoon-trend-research 01_research/audience-analyst.md
hook-analyst Reverse-engineering hooking/twist mechanisms webtoon-trend-research 01_research/hook-analyst.md
trend-synthesizer Comprehensive planning brief webtoon-trend-research 01_research/trend-brief.md
Scenario team concept-architect High concept/logline webtoon-scenario 02_story/concept.md
worldbuilder Worldbuilding·rules webtoon-scenario 02_story/world.md
character-designer Cast (appearance + personality + arc) webtoon-scenario 02_story/characters.md
series-plotter Series arc·episode map webtoon-scenario 02_story/series-arc.md
twist-master Per-episode twist design webtoon-scenario 02_story/twist-plan.md
tension-engineer Tension curve·cliffhanger webtoon-scenario 02_story/tension-curve.md
episode-outliner Episode beat sheet (50+ panels) webtoon-scenario 03_episode/ep{NN}_beatsheet.md
dialogue-writer Dialogue-driven script webtoon-scenario 03_episode/ep{NN}_script.md
script-editor Proofreading·twist clarity webtoon-scenario 03_episode/ep{NN}_script_final.md
Visual team art-director Style bible·consistency tokens·location tokens·speech bubble conventions webtoon-panel-breakdown 04_visual/style-bible.md, character-sheets.md
ref-sheet-artist Character multi-angle/expression reference sheets (preceding all panels) webtoon-panel-render 04_visual/refs/*.png, refs/INDEX.md
panel-director 50+ panel shot list (scene_id/location) webtoon-panel-breakdown 04_visual/ep{NN}_shotlist.md
letterer In-image speech bubble/dialogue bake spec webtoon-assembly 04_visual/ep{NN}_lettering.md
prompt-smith Per-panel codex prompt (bake + location + reference) webtoon-panel-render 04_visual/ep{NN}_prompts.md
panel-artist-a scene group A render webtoon-panel-render 05_panels/ep{NN}/panel_*.png
panel-artist-b scene group B render webtoon-panel-render 05_panels/ep{NN}/panel_*.png
panel-artist-c scene group C render webtoon-panel-render 05_panels/ep{NN}/panel_*.png
panel-validator Panel 6-axis validation·regeneration loop gate (general-purpose) webtoon-panel-render 04_visual/ep{NN}_validation.md
Assembly & review team episode-compositor Vertical scroll viewer assembly webtoon-assembly 06_assembly/ep{NN}/index.html
quality-reviewer QA review (general-purpose) webtoon-assembly 06_assembly/ep{NN}/qa_report.md
continuity-manager Inter-episode continuity webtoon-assembly 06_assembly/continuity.md
showrunner Oversight·sign-off·packaging webtoon-assembly RELEASE/ep{NN}/

Workflow

Phase 0: Context check (determine follow-up work)

_workspace/ Determine the execution mode based on existence and the user request.

  1. _workspace/ Does not exist → Initial run. Phase 1로.
  2. _workspace/ Exists + "next episode" request → New episode run. Increment {NN}, and 02_story·style-bible·character-sheets·refs/ (reference sheets)·continuity.md is reused (Read, no re-render — series consistency), only newly generated after 03_episode.
  3. _workspace/ exists + "only redo this episode's OO" request → partial re-run. Restructure and invoke only that stage's team, and overwrite only its outputs. For downstream stages (e.g., when the script is revised: shot list → render → assembly), re-run only as much as is affected.
  4. _workspace/ exists + new planning input → new planning run. Move to the existing _workspace/를 _workspace_{YYYYMMDD_HHMMSS}/and then Phase 1.

When partially re-running or running a new episode, include the previous output paths in the team members' prompts and instruct them to "Read first and reflect only the improvements."

Phase 1: Preparation

  1. Analyze user input — episode number {NN}, genre direction (if any), constraints (rating·length·tone).
  2. _workspace/00_input/brief.mdRecord input·episode number·constraints in .
  3. Ensure working directory: mkdir -p _workspace/{00_input,01_research,02_story,03_episode,04_visual,05_panels,06_assembly,RELEASE}.

Phase 2: Trend Research (Research Team)

Execution mode: agent team

  1. TeamCreate(team_name: "webtoon-research", members: [trend-scout, platform-ranker, audience-analyst, hook-analyst, trend-synthesizer]) — all agent_type=custom name, model: "opus".
  2. TaskCreateregister 5 tasks with . For the trend-synthesizer task, [coordinate] with the other 4 via depends_on.
  3. 4 researchers investigate in parallel and share interesting findings via SendMessage (hook-analyst↔trend-scout, etc.).
  4. trend-synthesizer synthesizes the 4 outputs → organize into 01_research/trend-brief.md.
  5. TeamDelete. (Outputs are preserved in _workspace/)

Phase 3: Scenario (Scenario Team)

Execution mode: agent team (pipeline + fan-out)

  1. TeamCreate(team_name: "webtoon-story", members: [concept-architect, worldbuilder, character-designer, series-plotter, twist-master, tension-engineer, episode-outliner, dialogue-writer, script-editor]).
  2. Task dependencies: concept → world → characters (sequential coordination) → series-arc → {twist-plan, tension-curve} (mutual coordination) → beatsheet → script → script_final.
  3. Guarantee a twist every episode: twist-master specifies a per-episode twist in twist-plan, and script-editor verifies in script_final that the twist is conveyed clearly.
  4. Secure 50+ panels' worth of content: episode-outliner splits the beats finely enough to produce 50+ panels' worth of content.
  5. Outputs: organize into 02_story/*, 03_episode/ep{NN}_script_final.md.
  6. TeamDelete.

Phase 4: Visual Production (Visual Team)

Execution mode: agent team (supervisor + generate-verify loop). Strictly enforce codex concurrent sessions ≤ 5.

  1. TeamCreate(team_name: "webtoon-visual", members: [art-director, ref-sheet-artist, panel-director, letterer, prompt-smith, panel-artist-a, panel-artist-b, panel-artist-c, panel-validator]).
  2. Art direction: art-director가 style-bible.md(art style·location token LOC_*·speech bubble visual conventions included) + character-sheets.md(consistency tokens + reference specs) and share with the team.
  3. Reference sheets first (consistency SSOT): ref-sheet-artist renders multi-angle/expression references of the main characters with codex (concurrent ≤5) → confirm in 04_visual/refs/→ hand off INDEX.md to prompt-smith·panel-validator. Panel rendering starts after references are confirmed. (For subsequent episodes, reuse the existing refs/.)
  4. Storyboard+lettering in parallel: panel-director가 ep{NN}_shotlist.md(50+ panels, each panel's scene_id/location), letterer works in parallel writing ep{NN}_lettering.md(in-image speech bubble bake spec, Korean kept short).
  5. Prompt synthesis: prompt-smith synthesized [style+location token+character token&reference anchor+composition/state color+speech bubble/Korean bake] ep{NN}_prompts.mdand distributed scene groups A/B/C. no text prohibited (speech bubbles must be drawn), negation no English/gibberish/misspelled text.
  6. Rendering (concurrency core): Render panels in batches of 5. Even with 3 panel-artists, total concurrent codex exec executions ≤ 5. The leader dispatches render passes sequentially or uses a single driver (scripts/codex_imagegen_batch.sh or episode render script) to auto-batch 5 at a time. 50 panels ≈ 10 waves.
  7. Validation-regeneration loop (core — until criteria are met): Immediately after batch rendering, panel-validator validates 6 axes (C1 character/reference, C2 background·location continuity, C3 speech bubble·Korean text, C4 prompt fidelity, C5 dialogue flow, C6 integrity·md5 duplication) → ACCEPT/REGEN decision. For REGEN panels, prompt-smith reinforces the prompt → the responsible panel-artist re-renders only that panel → re-validation. Maximum 3 times per panel, if exceeded, ACCEPT-FLAG (pass + record limit). When all panels pass, 04_visual/ep{NN}_validation.md complete.
  8. Deliverables: 04_visual/*(style-bible/character-sheets/refs/shotlist/lettering/prompts/validation), 05_panels/ep{NN}/panel_*.png(validation-passed version).
  9. TeamDeleteorganized as.

Phase 5: Assembly · Review (Assembly Review Team)

Execution mode: Agent team (generation-validation)

  1. TeamCreate(team_name: "webtoon-assembly", members: [episode-compositor, quality-reviewer, continuity-manager, showrunner]).
  2. episode-compositor assembles (speech-bubble-already-baked) panel PNGs into without overlay vertical scroll index.html(designing only spacing·rhythm).
  3. quality-reviewer performs progressive QA — first reads panel-validator's validation.md(identifying ACCEPT-FLAG), then reviews the entire episode for panel count 50+, reference appearance consistency, background continuity, baked Korean text accuracy/readability, dialogue flow, twist delivery, and damaged/duplicate images. PASS/FIX/REDO decision.
  4. FIX/REDO panels → re-inject into panel-validator loop (prompt-smith reinforcement + panel-artist re-render) or rework only that panel via SendMessage feedback (maximum 2 loops). If the twist is not revealed, feedback to the scenario team/letterer.
  5. continuity-manager가 continuity.mdupdate (inter-episode appearance/settings/hooks).
  6. After showrunner signs off, RELEASE/ep{NN}/final packaging + next episode seed (cliffhanger continuation) proposal.
  7. TeamDeleteorganized as.

Phase 6: Wrap-up

  1. _workspace/ fully preserved (intermediate deliverables = audit trail).
  2. Result summary to user: episode title·logline·one-line twist·panel count·viewer path (RELEASE/ep{NN}/index.html)·next episode seed.
  3. Feedback request: "Is there anything you'd like to fix in twist intensity/dialogue tone/art consistency?" (connects to Phase 7 evolution)

Data flow

trend-brief.md
   └→ concept → world → characters → series-arc → {twist-plan, tension-curve}
                                                       └→ beatsheet → script → script_final
script_final + characters
   └→ style-bible(+장소토큰 LOC_*, 말풍선 규약)/character-sheets
        └→ refs/*.png (레퍼런스 시트, 패널 전 선행)
        └→ shotlist(scene_id/location) + lettering(in-image 말풍선 명세)
              └→ prompts(스타일+장소+레퍼런스앵커+말풍선 베이크, scene A/B/C)
                    └→ panel_*.png ⇄ panel-validator 6축 검증-재생성 루프 ⇄ prompt-smith/panel-artist
                          └→ validation.md (전 패널 통과)
panel_*.png(말풍선 포함) → index.html(오버레이 없음) → qa_report → (재작업 루프) → RELEASE/ep{NN}/

Error handling

Situation Strategy
1 team member fails/stops Leader detects idle notification → SendMessage status check → restart or create replacement team member
Majority of team members fail Notify the user and confirm whether to proceed
codex render 0 bytes/corrupted Re-render only that panel (batch-wide rerender prohibited). After 2 failures, warn and pass, specify in report
Panel md5 duplicate (different panels have identical images) Delete the duplicate panel and re-render it individually. Size/header checks cannot catch this, so md5 checking is mandatory (actually occurred in EP01)
Abrupt background change (road→indoors, etc.) panel-validator C2 REGEN → prompt-smith strengthens location tokens (LOC_*) then re-renders only that panel
Korean speech bubble broken/typos panel-validator C3 REGEN → reinforce with shorter text, quotes, bold and re-render. After 3 failures, ACCEPT-FLAG the most accurate version + specify in report
Character appearance deviation panel-validator C1 REGEN → emphasize reference anchors and identifying markers then re-render
Entire codex is slow (450 seconds+) Suspect plan limit → codex login status check, reduce concurrency and retry
Fewer than 50 panels Request episode-outliner/panel-director to split and add beats
Twist unclear (QA REDO) Give feedback to scenario team (twist-master/script-editor) or letterer then rework
Data conflict Cite sources, deletion prohibited
Infinite rework Maximum 3 regenerations per panel, maximum 2 step-wise rework loops. If exceeded, proceed with current state and report the limits

Test scenario

Normal flow

  1. User: "Make episode 1 of a webtoon reflecting trends."
  2. Phase 1: record brief, _workspace/ create.
  3. Phase 2: 5 research team members → trend-brief.md.
  4. Phase 3: 9 scenario team members → script_final.md (1+ twists, 50+ panels worth of content).
  5. Phase 4: 9 visual team members → reference sheets first → 50+ panel PNGs (speech bubbles baked in, batches of 5, concurrency ≤5) → panel-validator 6-axis validation-regeneration loop passes all panels → validation.md.
  6. Phase 5: 4 assembly/inspection team members → index.html without overlays + qa PASS → RELEASE/ep01/.
  7. Result: RELEASE/ep01/index.html vertical-scroll webtoon (including speech bubbles) + next episode seed.

Error flow (abrupt background change + Korean text corruption regeneration)

  1. After Phase 4 rendering, panel_023's background is the same scene (road) but comes out as indoors, and panel_031's speech bubble Korean text is corrupted.
  2. panel-validator judges C2 (023) and C3 (031) as REGEN → records reasons and correction instructions in validation.md.
  3. prompt-smith strengthens location tokens (LOC_*) for 023 and reinforces text with quotes, bold, shorter for 031 → the responsible panel-artist re-renders only those 2.
  4. panel-validator re-validates: 023 ACCEPT, 031 still has some shaky characters even after 3 attempts → ACCEPT-FLAG the most accurate version + specify in report.
  5. After all panels pass, proceed with assembly, specify flagged panel limitations in qa_report and RELEASE notes.

Follow-up work guide

  • "Next episode" → Phase 0 case 2 (new episode). Reuse worldbuilding/style/continuity, only new content after 03_episode.
  • "Make the twist stronger" → Partial re-run of the scenario team (twist-master, script-editor) → re-run affected downstream stages.
  • "Redo panel N" → Partial re-run of the visual team (prompt-smith reinforcement + re-render by the relevant panel-artist + re-validation by panel-validator) → reassembly.
  • If the same type of feedback repeats 2+ times, propose improvements to the relevant skill/agent definition (harness evolution). Record changes in the CLAUDE.md change history.