Goal: Produce an original Hangul font the world has never seen through a sample → feedback → improvement loop.
Team Composition (12 agents, definition: .claude/agents/)
| Agent | Role | Primary Skill |
|---|---|---|
| font-director | Creative director, approve/reject | — |
| hangul-analyst | Jamo·block·coverage analysis | hangul-composition |
| concept-artist | Concept discovery + sample glyphs | hangul-composition, glyph-design |
| choseong-designer | 19 initial consonants + variants | glyph-design |
| jungseong-designer | 21 medial vowels + variants | glyph-design |
| jongseong-designer | 27 final consonant types | glyph-design |
| latin-symbol-designer | Latin·numbers·symbols | glyph-design |
| composition-engineer | Slot table, syllable composition | hangul-composition, glyph-design |
| specimen-designer | Specimen production·distribution | font-specimen |
| font-engineer | TTF build | font-build |
| type-qa | Consistency·boundary verification | glyph-design, font-build |
| feedback-analyst | Feedback → change spec | design-feedback |
In every Agent call, specify model: "opus".
Phase 0: Context Check (required every run)
_workspace/ Determine the execution mode based on the status:
_workspace/None → Initial run: Full run from Phase 1_workspace/Present + feedback/partial revision request → Partial rerun: Start from feedback-analyst and re-invoke only the agents specified by the change spec. First read the_workspace/feedback/·_workspace/decisions/history. If it is a request to modify specific characters·jamo, go directly to the jamo revision loop (12) in Phase 3 — modify jamo definitions → regenerate → rebuild_workspace/Present + completely new direction request → New run: Move the existing to_workspace_prev/and then initial run- If it is ambiguous which one it is, confirm with the user — do not arbitrarily decide to overwrite existing work
Workflow
Phase 1: Concept Exploration — Execution mode: sub-agents (parallel)
Since this is independent exploration, team communication is unnecessary. To prevent contamination between concepts, they should rather not know about each other.
- Generate sample syllable set·jamo decomposition table with hangul-analyst (
_workspace/analysis/) - Spawn 2–4 concept-artists in parallel with briefs on different axes (structure/stroke/space) (
run_in_background: true)- Each produces concept.md + sample glyph SVG in
_workspace/concepts/{콘셉트명}/Produce concept.md + sample glyph SVG
- Each produces concept.md + sample glyph SVG in
- font-director evaluates the results (
_workspace/decisions/round-1.md)
Phase 2: Specimen Production and Feedback — Execution mode: sub-agents (pipeline)
- type-qa verifies glyph conventions (no stroke, viewBox, etc. — mechanical detection first)
- specimen-designer가
_workspace/specimens/round-{N}.htmlproduction + Artifact distribution - Await user feedback — receive answers to the question block in the specimen
- feedback-analyst converts feedback into a change spec (
_workspace/feedback/round-{N}.md) - Recall the responsible agent according to the changed spec → return to 5. Repeat this loop until the concept is finalized
Phase 3: Main production — Jamo combination system (after concept finalization) — Execution mode: Hybrid
Main production is not manual syllable work but Jamo definitions × combination engine(built in Round 4, see the "Jamo combination system" section of the hangul-composition skill). Cheonjiin system: _workspace/glyphs/천지인/system/ — owned by composition-engineer.
- System construction/extension (if new concept): composition-engineer verifies the context box (contexts.json) and engine — visual comparison between handcrafted samples and engine-regenerated output is the acceptance criterion
- Write Jamo definitions (sub-parallel): choseong/jungseong/jongseong-designer each handle jamo_cho/jung/jong.json — files are separate so there is no conflict. Each generates test syllables with compose.py → renders a contact sheet → submits after passing the Read-based inspection loop
- Full generation + sample QA:
compose.py --ks2350to generate 11,172 characters → visual verification of samples including everyday words + extreme combinations (double consonants, compound vowels, double final consonants) - Jamo revision loop (when there is feedback/QA criticism): do not fix individual SVGs, but fix the cause in the Jamo definition → regenerate everything → re-verify samples. Reason: individual fixes are lost on regeneration, whereas a Jamo fix applies uniformly to all characters that use that Jamo
- Update the draft (the 5–8 loop of Phase 2) — do not increase coverage without feedback
Phase 4: Build·release — Execution mode: Sub-agent
- Produce Latin and symbols appearing in the draft with latin-symbol-designer (when requested)
- font-engineer builds the TTF —
--family는 ASCII("Cheonjiin") +--family-ko "천지인"(font-build skill: non-ASCII default names are rejected on macOS installation) + build report - type-qa verifies the build artifact boundaries (manifest ↔ TTF glyph count·cmap cross-check) + confirms actual typesetting rendering with PIL
- When release is requested: the release procedure of the font-build skill —
fonts/keep only the latest version, update the README version table, push with a Korean commit message (github.com/revfactory/font-harness)
Data transfer protocol
- File-based (primary): all artifacts are stored at
_workspace/convention paths. Filenames and paths are the interface - Team mode: TaskCreate (tasks·dependencies) + SendMessage (real-time coordination such as interface coordinates and rejection reasons)
- Sub mode: return values (summary·warnings) + files (artifacts)
_workspace/
├── analysis/ # hangul-analyst: 음절 분해, 커버리지, 제작 브리프
├── concepts/{명}/ # concept.md + glyphs/*.svg + manifest.json
├── glyphs/{명}/
│ ├── syllables/ # 손 제작 음절 (초기 라운드 검증용)
│ ├── system/ # 자모 조합 시스템: jamo_*.json + contexts.json + compose.py
│ └── generated/ # 엔진 생성 전체 음절 + manifest.json (재생성 가능 — 수정 금지)
├── specimens/ # round-{NN}.html (+ round-{N}-copy.json)
├── feedback/ # round-{N}.md
├── decisions/ # font-director 결정 기록
├── qa/ # 검수 리포트
└── build/ # TTF (버전별 보존)
fonts/ # 배포 폴더 — 최신 릴리스 1개만 (README·GitHub 공개용)
Error handling
| Situation | Handling |
|---|---|
| Agent artifact missing/convention violation | Retry once (specifying the violation) → if it fails again, proceed excluding that item and state the omission in the draft·report |
| 1 failure during parallel concept exploration | Proceed with the rest, record the failed concept in the decision document |
| Slot/spec conflict (contradiction between designers) | Do not delete; list both → font-director arbitration, record decision |
| Build-failing glyph | Exclude from build + state in report, fix as top priority next round |
| Ambiguous feedback interpretation | Do not proceed arbitrarily; ask the user with the minimum number of clarifying questions |
Test scenarios
Normal flow (initial run): "Make an original Korean font sample" → Phase 0 (initial assessment) → analysis → 3 concepts in parallel → QA → distribute draft round-01 → present feedback questions. Success criteria: _workspace/concepts/ 2 or more, draft contains type labels, explicit omissions, and a question block.
Error flow (partial rerun + artifact defect): "The final consonant is too big, fix it" → Phase 0 (partial rerun assessment) → feedback-analyst generalizes to "adjust type 4·5·6 final-consonant slots" → re-invoke composition-engineer → during recomposition, discover a ㅄ glyph convention violation → return once and retry → on failure, update the draft with a placeholder for that syllable + state the omission. Success criteria: fix by rule unit rather than only the pointed-out character, and the omission is marked in the draft.