| name | webtoon-harness |
| description | Run an end-to-end webtoon production harness — a 27-agent / 4-phase Claude Code agent team that takes a single episode from trend research to a finished vertical-scroll viewer. Use when the user wants to produce a webtoon episode: research popular webtoon trends, write a dialogue-heavy, high-tension, twist-every-episode scenario, render character reference sheets first, render 50+ panels per episode with in-image speech-bubble (Korean dialogue) baking via codex-image, run a generation–validation loop, and assemble a vertical-scroll viewer. Triggers on: webtoon harness, make a webtoon, produce a webtoon episode, webtoon scenario to image, next episode, redraw panel, stronger twist, webtoon trend research. Web data gathering for the trend-research phase routes through the `scrapling` skill. Pure webtoon recommendation/critique is answered directly.
|
| allowed-tools | Bash Read Write Edit Glob Grep WebFetch Agent |
| compatibility | Requires Claude Code with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 for agent-team execution, and the codex CLI (ChatGPT OAuth) for parallel panel image rendering (max 5 concurrent codex sessions). Trend-research web extraction uses the scrapling skill (Python 3.10+). |
| license | MIT |
| metadata | {"tags":"webtoon, agent-team, multi-agent, content-production, codex-image, panel-render, trend-research, scrapling, vertical-scroll, claude-code","version":"1.0","source":"https://github.com/revfactory/webtoon-harness"} |
webtoon-harness — Webtoon Production Agent Team
Keyword: webtoon harness · make a webtoon · produce a webtoon episode · next episode · redraw panel
Respect every source site's terms, robots.txt, rate limits, and copyright when gathering trend
data, and respect platform/IP rules when publishing generated webtoons.
webtoon-harness packages the revfactory/webtoon-harness
Claude Code harness as an installable jeo-skills plugin. It builds one webtoon episode end to end
with 27 specialized agents organized into 4 phase-rebuilt teams: research popular trends →
write a dialogue-driven, high-tension, twist-every-episode scenario → render character reference
sheets first → render 50+ panels per episode with in-image speech-bubble baking → run a
generation–validation loop until every panel passes → assemble a vertical-scroll viewer.
Agent teams require CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1.
When to use this skill
- Produce a full webtoon episode from trend research to a finished vertical-scroll viewer
- Research webtoon trends: genre/trope momentum, platform rankings, reader reactions, hook/twist reverse-engineering
- Write a dialogue-heavy, high-tension scenario with a twist every episode and a 50+ panel beat sheet
- Render character reference sheets first (the single source of truth for cross-episode consistency)
- Batch-render 50+ panels with in-image Korean speech bubbles baked in, then validate and regenerate
- Assemble panels into a no-overlay vertical-scroll viewer and package a release
- Follow-ups: "next episode", "stronger twist", "redraw panel 23", "re-run only the visual phase"
For pure recommendation or critique ("what should I read?"), answer directly — do not run the harness.
Install (plugin)
npx skills add https://github.com/akillness/jeo-skills --skill webtoon-harness
bash scripts/install.sh
TARGET=/path/to/project bash scripts/install.sh
GLOBAL=1 bash scripts/install.sh
The installer is idempotent. It clones revfactory/webtoon-harness and copies its .claude/agents/
and .claude/skills/ into the target project so Claude Code can discover the agents and the
webtoon-orchestrator entry-point skill. See references/install.md.
Execution mode: phase-rebuilt agent teams
Only one team is active per session. For each phase, create a team with TeamCreate, run it, then
TeamDelete before building the next phase's team. Each team leaves its artifacts under
_workspace/ so the next team can pick them up with Read. Spawn every team member and Agent
call with model: "opus".
The 27 agents (4 teams)
| Team | Members | Role |
|---|
| Research | trend-scout, platform-ranker, audience-analyst, hook-analyst, trend-synthesizer | Genre/trope momentum · platform rankings & serialization structure · reader reactions/drop-off · hook/twist reverse-engineering → one planning brief |
| Scenario | concept-architect, worldbuilder, character-designer, series-plotter, twist-master, tension-engineer, episode-outliner, dialogue-writer, script-editor | High-concept · worldbuilding · cast · series arc · twist-per-episode · tension curve · 50+ panel beat sheet · dialogue-driven script · edit pass |
| Visual | art-director, ref-sheet-artist, panel-director, letterer, prompt-smith, panel-artist-a/b/c, panel-validator | Style bible & consistency/location/bubble tokens · reference sheets · shot list · lettering · prompt synthesis · parallel panel render · 6-axis validation loop |
| Assembly & QA | episode-compositor, quality-reviewer, continuity-manager, showrunner | Vertical-scroll viewer assembly · QA · continuity tracking · sign-off & release packaging |
Full agent roster and outputs → references/agent-teams.md.
Workflow — 6 phases
Phase 0 Context check — new series vs next episode; reuse world/style/continuity
Phase 1 Prep — record brief, create _workspace/
Phase 2 Research — 4 parallel researchers → trend-brief.md (web data via scrapling)
Phase 3 Scenario — high-concept → world → cast → arc → {twist, tension} → beatsheet → script_final
Phase 4 Visual — style bible → reference sheets FIRST → shotlist+lettering → prompts
→ 50+ panels (codex-image, batches of 5, ≤5 concurrent)
⇄ panel-validator 6-axis validate–regenerate loop → validation.md
Phase 5 Assembly & QA — no-overlay index.html (bubbles baked into panels) → qa PASS → RELEASE/ep{NN}/
Phase 6 Wrap-up — continuity update, next-episode seed
Detailed dataflow and per-phase outputs → references/workflow.md.
Trend research via scrapling (Phase 2)
The research team gathers live data — platform rankings, serialization structure, comment-reaction
patterns — and that web extraction routes through the scrapling skill, not ad-hoc browser
defaults. Pick the lightest workable Scrapling mode (parser-only → HTTP fetch → JS browser → stealth)
and respect each site's terms, robots.txt, rate limits, and copyright. Attach a source URL and
observation date to every claim; when sources conflict, record both rather than deleting one.
See references/trend-research-scrapling.md.
Design principles
- Dialogue-driven, high-tension, twist-every-episode — minimize narration; carry tension, info, and reversals through character dialogue and action.
- 50+ panels per episode — split beats finely to match the vertical-scroll rhythm.
- Consistency first — inject reference sheet → consistency token → location token into every prompt; the validator catches md5 duplicates, abrupt background swaps, and broken Korean text.
- In-image bubble baking — speech bubbles and Korean dialogue are rendered into the panel PNG, so assembly is a pure ordered image strip with no text overlay (no double captioning).
- Audit trail — preserve every intermediate artifact under
_workspace/; never delete it.
Follow-up routing
- "next episode" → Phase 0 case 2: reuse world/style/continuity, regenerate from
03_episode onward.
- "stronger twist" → partial Scenario re-run (twist-master, script-editor) → re-run affected downstream phases.
- "redraw panel N" → partial Visual re-run (prompt-smith reinforce + that panel-artist re-render + panel-validator re-check) → re-assemble.
- Repeated feedback of the same kind (2×+) → propose improving the relevant skill/agent definition (harness evolution); log the change.
Limits & error handling
| Situation | Handling |
|---|
| codex global slowdown (450s+) | Suspect plan limits → check codex login status, reduce concurrency, retry |
| Character drift | panel-validator C1 REGEN → reinforce reference anchor / identity marks, re-render |
| Broken Korean text | prompt-smith reinforces quotes/bold/short text → re-render; if still shaky after 3 tries, ACCEPT-FLAG the best version and note it |
| < 50 panels | Ask episode-outliner/panel-director to split more beats |
| Unclear twist (QA REDO) | Feed back to twist-master/script-editor or letterer, re-work |
| Infinite rework | Max 3 regenerations per panel; max 2 rework loops per phase. On overflow, proceed with current state and report the limit |
References