Portfolio Manage — the portfolio writing flow, material to a reviewed deck
The user's one entry for writing a portfolio, first draft to final deck; it runs the other portfolio
skills as its steps. Not for a one-off review of someone else's document (feedback) or a deck
that is not a portfolio (deck-builder).
Engine, run from the workspace root: node ${CLAUDE_PLUGIN_ROOT}/skills/manage/scripts/manage.mjs <cmd>
(layout, scaffold, selection, snapshot (--stable), renumber, verify, sync, verify-build, compare,
claims, settle, intake, step, round, round-check, questions, keep-render; one JSON object each). A real run's mismatch becomes a fixture test first.
Workspace
CLAUDE.md the guide every session loads: these names, `scaffold` first (written by scaffold if none)
portfolio-skeleton.md the portfolio design: every wording change starts here (`version: vX.Y.Z` at the top)
portfolio-deck.mdx its build source (output: build/deck.pptx)
resume-skeleton.md the resume design (step 9), facts from portfolio-skeleton; resume-deck.mdx its build source
template/ template.pptx, resume template (.docx), NOTES.md (quirks), analysis.md (step 5),
next-render.md (step 7: render problems held for the next render)
deck.catalog.md slot ids and character budgets — regenerated only (deck-builder catalog)
build/ deck.pptx, deck.pdf, resume.docx, resume.pdf, pages/ (previews) — regenerated in full
snapshot/ skeleton-vX.Y.Z/, deck/deck-vX.Y.Z/, scores/, render/deck-vX.Y.Z/ (pages/, 검토.md),
README.md (`## deck` and `## skeleton` tables); rounds/round-NN/ = one round's record
stable/ YYYY-MM-DD-skeleton-vX.Y.Z/ frozen baseline (skeleton, deck + resume source, style), README.md row; never edited
style/ 문체.md 표기규칙.md 레이아웃.md 이력서-서식.md — the user's fixed rules (`scaffold` templates)
consensus/ decisions/ debate transcripts kept unedited (`<date>-<topic>.md`); decisions `NN-<topic>.md`
analysis/ link to the user's raw archive outside the workspace: read-only
feedback/ YYYY-MM-DD.md — one `feedback` review per day, ending with its fix table
sources/ material: material/<id>/ items (`material` skill), or a ledger NN_프로젝트/A1….md read as it is;
선정.md (step 2: 글감 · 쓰임 본문|부록|제외 · 쪽 · 이유 · 대표 — the one place an exclusion lives)
docs/확정사실.md settled facts, corrections; `## 서술 기준` and `## 구조` lines no check raises again — read before any work
docs/확인할-것.md questions a round could not settle: `questions --add` each round, answered ones read back (`--read`)
_archive/ old plans and deck backups, moved here, never deleted
These names are the contract: a new file takes its name from this tree; older names are legacy (layout),
renamed by scaffold. Root holds only CLAUDE.md and the skeleton and deck pairs; tidying moves to _archive/
(references/workspace.md). Page headings: ## N <title> or ## N쪽 · <title>, one per page
(appendix too); references N쪽, N·M쪽, N~M쪽 and a 쪽 table column. A # heading bounds a
block that moves with its pages. History lines (vX.Y.Z 변경…, vX.Y.Z(…, >-quoted) stay as written.
Process
Nine steps, in order; a round may start again at 2 or 4 (step 8 says where); 9 once the deck is done.
Steps 1-7 run as one harness-shaped pass, n times (n from the user, default n = 2): plan the round
→ criteria per step (step) → do → check (a command, else a judge that is not the actor) → gate
(step) → round. A round is the full flow once over the whole deck, not a part of it. feedback runs every
round after step 7 from the main context, intake takes its findings; the loop stops at round's continue: false.
A step that could not run is skipped with its reason, a gap, never pass. Mechanics: references/rounds.md.
The round's record (round writes snapshot/rounds/round-NN/; detail references/workspace.md): that round's
skeleton, feedback (revision mode: previous vs this version) and scores, pattern two-version scores, compare,
noah.md, overrides.md, report.md. noah (the user's proxy) gives its verdict every round, written apart from the panel.
Every edit in report.md carries its source (feedback priority | panel | noah), the evidence file and the reason; a
finding not used says why; "사용자가 정할 구조" lists what only the user decides. Run round-check --round N before
reporting: an edit without a reason is a problem. Each question the round could not settle goes to
questions --add --text … --page N; at the start questions --read returns the answered ones: use them, do not ask again.
A baseline is frozen with snapshot --stable --note … (rounds compare against it, it is never edited).
Which gates block. Structure only: the order of pages and what goes in or out (step 2's table, step 3's
blocks, a step 4 order or page-count change). It is asked, answered by the user, else noah, else the
proposal (answered_by), and is the only wait; the round never stops on it. Everything else goes ahead on
the proposal and is reported under "기본값으로 정한 것": wording, fixes from feedback, the deck source,
the snapshot. A round always builds (the user's standing word: "결과물도 검증하고, 더 빌드해보자"):
sync, deck-builder build, verify-build; render only when the user says.
- Open, every session, unasked.
scaffold (renames old names, creates only what is missing,
never overwrites, appends the guide once to a CLAUDE.md that lacks it, creates style/ templates and docs/확인할-것.md;
its renamed/created/appended in one report line), then layout (its warnings: a broken analysis/ link, an
engine path in CLAUDE.md or style/ that is not on disk), docs/확정사실.md, questions --read and the "맞춰야 할 것" table. A stray root file (*.conflict-*, a scratch script) is named and moved
to _archive/ once the user agrees.
- Collect. New material → run
material from the workspace root (items land in
sources/material/). Evidence order: sources/ → the user's analysis notes they point to →
the user. Search before asking; ask only what only the user can answer; everything else stays
[확인 필요: …].
- Select. Before any page is written, one row per item in
sources/선정.md: 글감 · 쓰임
(본문 | 부록 | 제외) · 쪽 · 이유 · 대표 (the one problem, decision or result the item leads
with, in the material's words). Picked by signal (references/principles.md § Composition);
propose the table (structure: asked). Mark each item material log --id <id> --note "선정: <쓰임>" --status selected|excluded --later-step manage. An exclusion lives only here, not
in 확정사실.md. selection reads the table back; a 쓰임 outside the three is a problem.
- Draft. No skeleton yet, or a new project → write it from the
본문/부록 items only: one
project as why → the user's role and decision → result; strengths chosen from the evidence.
Each page that carries a figure or a reason cites its items on one line under its heading, 글감: <id>, <id> (a ledger item is
NN_프로젝트/A1); a reason the item does not give is written [확인 필요: …], never filled in.
The structure (blocks, order, body vs appendix) is asked; wording is not. A written
item is marked material log --id <id> --note "skeleton <N>쪽" --status written --later-step manage;
in a ledger workspace (material ledger lists items) material ledger-log --item <NN_프로젝트/A1>
with the same flags. The ledger is never moved or rewritten.
- Revise, from several viewpoints.
claims first (each problem is an edit; 본인 확인 stays),
then recruiter/engineer (feedback, pattern) → interviewer (mock-interview question-list
mode) → the user's own bar (noah), each viewpoint's findings through settle. snapshot before
the round's first edit (--kind skeleton|deck|resume); carried wording passes the decision gate (확정사실.md, 선정.md);
an order change is one renumber; the align table stays at the end of portfolio-skeleton.
Rules and the order of each: references/revise.md.
- Template. Once per template (again when it changes):
deck-builder catalog, then
scaffold copies its Palette, Type and Voice into template/analysis.md; add the style rules the
pages must keep (spacing, alignment, emphasis — only what the template shows), the safe area
(the margins nothing crosses) and the figure colours. A figure drawn with write:diagram takes
its colours from ## 그림 색.
- Deck. Carry portfolio-skeleton into portfolio-deck.mdx:
sync lists each skeleton slot missing or
different in the deck source, sync --write carries them (never a slide); against deck.catalog.md budgets and
template/analysis.md, then deck-builder check (an untouched chart slot is warned: its numbers are the template's
sample) and claims --file portfolio-deck.mdx (each slide cites in its comment: ## @s3 2쪽 · 글감: <id>). Fine style changes go through the template's run formats, never by hand in the pptx.
snapshot --kind deck --level … --note only when asked (the note names its skeleton).
- Build, render, review. Every round builds, then
verify-build (each skeleton slot's text is in
build/deck.pptx, footnote marks and picture files exist). render only when the user says. After a render run
render-review (deck and resume alike): keep-render, then every page against safe area,
readability, AI traces and the skeleton, one 검토.md row per page. A problem seen there or in
PowerPoint goes into template/next-render.md (쪽 · 칸 · 증상 · 확인 · 적용), all applied before
the next render; fixed at once only when the user says ("고쳐", "맞춰줘"). A frame the deck cannot
move is fixed in the template only on the user's word (references/template-edit.md).
- Feedback on the deck.
feedback on the rendered PDF with the material and 확정사실 in reach,
saved to feedback/YYYY-MM-DD.md with its fix table. Each finding goes back to material: a claim
the deck cannot defend → its item's gaps.md; a result the user restores → result.md labelled
본인 확인. The next round starts at 2 (selection changed) or 4 (wording).
- Resume.
resume-skeleton.md (the template's section order, facts from portfolio-skeleton
only) → resume-deck.mdx → claims --file resume-deck.mdx → resume-builder (catalog, build into
the user's template, final checks, render when told, fit by over_lines) → render-review (step 7).
The user's material never goes to git: the engine refuses a workspace git would carry.
Always, portfolio and resume alike (user, 2026-11-10: "safe area가 있는거 시인성, 가독성이 좋은거
like-me로 ai티를 줄이는거 이건 포폴, 이력서 공통임"):
- Safe area, readable. Nothing crosses the margins (
layout's safe_area, from style/레이아웃.md, is what
deck-builder uses) or touches another box; type and spacing stay the template's, one idea per line, no lone last word. Cut text, never shrink type.
- In the user's voice. Read
style/문체.md and style/표기규칙.md before any rewrite; every sentence written into a
skeleton or deck source goes through write:like-me (samples: style/문체.md's own paragraphs and what the user wrote
alone in sources/, never AI-edited text) to strip machine tells; then claims. A resume follows style/이력서-서식.md.
- No traces. Deliverables carry no AI or tool trace; the builders' final check catches it.
- Corrections. When the user corrects a fact, in the same turn: append to the material file
## 정정 — 본인 확인 YYYY-MM-DD with their words verbatim; add one line to 확정사실.md;
noah_collect the words when noah is connected. The user's word beats the record.
- Did it read better. After a round's edits, score the snapshot and the new skeleton with the same rubric, then
compare (lost_ground = worse): references/rounds.md. A vague answer ("바꾸고"): do the shared part, hold the rest, say so.
Output Template
[결론 한 줄 — 무엇을 바꿨고 무엇이 남았나] (단계 N · skeleton vA → vB)
바꾼 것
- N쪽 칸: "현재 문구" → "바뀐 문구" (한 줄씩, 자리만 적지 않는다)
기본값으로 정한 것 (다르면 말씀해 주세요)
- …
정하실 것
1. 본인만 답할 수 있는 것만
이번 라운드 <나아짐 | 나빠짐 | 섞임 | 그대로> — 항목: 이전 → 지금 (떨어진 항목은 이전 문구와 함께)
확인한 것 (템플릿이나 원재료를 고쳤을 때: 백업 위치, check 결과)
claims: 문제 <n> (근거 없음 n · 글감에 없음 n · 확인 필요 n · 인용 없음 n) · 본인 확인 <n> · 덱에 없는 성과 <n>
renumber/verify: <changes, warnings, problems | 안 함> · sync: <빠진 n · 다른 n> · deck check: <warnings | 안 함> · build: <함> · verify-build: <빠진 문구 n · 각주 n · 그림 파일 n> · render: <함 | 안 함>
라운드: <n>회 중 <k>회 · 종료 <stop_reason> · 게이트 gap <n> (건너뛴 단계: 이유) · 패널 <4인 | 단일> · 심사 <subagent | in-context>
기록: round-NN/ · round-check 문제 <n> · noah 판정 <값> · 열린 질문 <n> · 기준본 <stable 폴더 | 없음>
Korean, conclusion first, plain headings, no bold runs, no invented names or labels; a number says what it counts.
Not built → the built files are older. A remark read one way → say the reading, ask. Held is said as held.
What Claude Does / What You Do
| Claude |
You |
| Runs material, feedback, pattern, rewrite and deck-builder as steps of one loop |
Hand over material; answer only what only you know |
| Proposes the selection table (본문/부록/제외 · 대표); structure is the only wait |
Approve what goes in and what each item leads with |
| Snapshots before editing, renumbers in one pass, verifies references |
Approve structure changes (page count, blocks, order) |
| Runs a whole round: sync, build, verify-build; wording goes ahead and is reported |
Correct facts in your own words; say when to render |
Related Skills
material — step 1, collects material into sources/material/; its status follows steps 2-3
feedback, pattern, rewrite, mock-interview — steps 4 and 8; write:like-me — the voice
deck-builder — steps 5-7; resume-builder — step 9 (docx build, PDF, fit); render-review — step 7
and the resume's render; write:diagram — figures; job-application-workflow — the whole application
references/principles.md — principles for any candidate and the mistakes this flow stops
1---2name: manage3effort: high4description: >-5 Use when writing a developer portfolio: "포트폴리오 쓰자", "포트폴리오 작업하자", "피드백 돌려줘",6 "순서 바꾸자", "스냅샷 찍고". The whole writing flow — material, write, feedback, deck —7 keeping versions, corrections and settled decisions.8scenarios:9 - "Let's keep working on my portfolio deck — run feedback on the skeleton and fix what it finds."10 - "Swap the two pages of the search block and fix every page reference."11 - "That's wrong, the retry policy was theirs, not mine — record it."12 - "포트폴리오 작업 이어서 하자. 피드백 돌리고 고쳐줘"13 - "검색 블록 두 장 순서 바꾸고 쪽 번호 다 맞춰줘"14 - "스냅샷 찍고 portfolio-deck.mdx에 반영해서 check까지만 해줘"15compatibility:16 required:17 - node 18+ (stdlib only) for scripts/manage.mjs18 optional:19 - noah MCP (noah_collect) — the user's corrections are sent there too when it is connected20 remote_mcp_note: >-21 noah가 없으면 정정은 글감 파일과 docs/확정사실.md에만 남고, 그 사실을 한 번 알립니다.22metadata:23 source: https://skillmd.com/skills/newkayak12/manage24---2526# Portfolio Manage — the portfolio writing flow, material to a reviewed deck2728The user's one entry for writing a portfolio, first draft to final deck; it runs the other portfolio29skills as its steps. **Not for** a one-off review of someone else's document (`feedback`) or a deck30that is not a portfolio (`deck-builder`).3132Engine, run from the workspace root: `node ${CLAUDE_PLUGIN_ROOT}/skills/manage/scripts/manage.mjs <cmd>`33(`layout`, `scaffold`, `selection`, `snapshot` (`--stable`), `renumber`, `verify`, `sync`, `verify-build`, `compare`,34`claims`, `settle`, `intake`, `step`, `round`, `round-check`, `questions`, `keep-render`; one JSON object each). A real run's mismatch becomes a fixture test first.3536## Workspace3738```39CLAUDE.md the guide every session loads: these names, `scaffold` first (written by scaffold if none)40portfolio-skeleton.md the portfolio design: every wording change starts here (`version: vX.Y.Z` at the top)41portfolio-deck.mdx its build source (output: build/deck.pptx)42resume-skeleton.md the resume design (step 9), facts from portfolio-skeleton; resume-deck.mdx its build source43template/ template.pptx, resume template (.docx), NOTES.md (quirks), analysis.md (step 5),44 next-render.md (step 7: render problems held for the next render)45deck.catalog.md slot ids and character budgets — regenerated only (deck-builder catalog)46build/ deck.pptx, deck.pdf, resume.docx, resume.pdf, pages/ (previews) — regenerated in full47snapshot/ skeleton-vX.Y.Z/, deck/deck-vX.Y.Z/, scores/, render/deck-vX.Y.Z/ (pages/, 검토.md),48 README.md (`## deck` and `## skeleton` tables); rounds/round-NN/ = one round's record49stable/ YYYY-MM-DD-skeleton-vX.Y.Z/ frozen baseline (skeleton, deck + resume source, style), README.md row; never edited50style/ 문체.md 표기규칙.md 레이아웃.md 이력서-서식.md — the user's fixed rules (`scaffold` templates)51consensus/ decisions/ debate transcripts kept unedited (`<date>-<topic>.md`); decisions `NN-<topic>.md`52analysis/ link to the user's raw archive outside the workspace: read-only53feedback/ YYYY-MM-DD.md — one `feedback` review per day, ending with its fix table54sources/ material: material/<id>/ items (`material` skill), or a ledger NN_프로젝트/A1….md read as it is;55 선정.md (step 2: 글감 · 쓰임 본문|부록|제외 · 쪽 · 이유 · 대표 — the one place an exclusion lives)56docs/확정사실.md settled facts, corrections; `## 서술 기준` and `## 구조` lines no check raises again — read before any work57docs/확인할-것.md questions a round could not settle: `questions --add` each round, answered ones read back (`--read`)58_archive/ old plans and deck backups, moved here, never deleted59```6061These names are the contract: a new file takes its name from this tree; older names are `legacy` (`layout`),62renamed by `scaffold`. Root holds only CLAUDE.md and the skeleton and deck pairs; tidying moves to `_archive/`63(`references/workspace.md`). Page headings: `## N <title>` or `## N쪽 · <title>`, one per page64(appendix too); references `N쪽`, `N·M쪽`, `N~M쪽` and a `쪽` table column. A `# ` heading bounds a65block that moves with its pages. History lines (`vX.Y.Z 변경…`, `vX.Y.Z(…`, `>`-quoted) stay as written.6667## Process6869Nine steps, in order; a round may start again at 2 or 4 (step 8 says where); 9 once the deck is done.70Steps 1-7 run as one harness-shaped pass, n times (n from the user, default n = 2): plan the round71→ criteria per step (`step`) → do → check (a command, else a judge that is not the actor) → gate72(`step`) → `round`. A round is the full flow once over the whole deck, not a part of it. `feedback` runs every73round after step 7 from the main context, `intake` takes its findings; the loop stops at `round`'s `continue: false`.74A step that could not run is `skipped` with its reason, a gap, never pass. Mechanics: `references/rounds.md`.7576**The round's record** (`round` writes `snapshot/rounds/round-NN/`; detail `references/workspace.md`): that round's77skeleton, feedback (revision mode: previous vs this version) and scores, pattern two-version scores, `compare`,78`noah.md`, `overrides.md`, `report.md`. noah (the user's proxy) gives its verdict every round, written apart from the panel.79Every edit in `report.md` carries its source (feedback priority | panel | noah), the evidence file and the reason; a80finding not used says why; "사용자가 정할 구조" lists what only the user decides. Run `round-check --round N` before81reporting: an edit without a reason is a problem. Each question the round could not settle goes to82`questions --add --text … --page N`; at the start `questions --read` returns the answered ones: use them, do not ask again.83A baseline is frozen with `snapshot --stable --note …` (rounds compare against it, it is never edited).8485**Which gates block.** Structure only: the order of pages and what goes in or out (step 2's table, step 3's86blocks, a step 4 order or page-count change). It is asked, answered by the user, else noah, else the87proposal (`answered_by`), and is the only wait; the round never stops on it. Everything else goes ahead on88the proposal and is reported under "기본값으로 정한 것": wording, fixes from feedback, the deck source,89the snapshot. A round always builds (the user's standing word: "결과물도 검증하고, 더 빌드해보자"):90`sync`, `deck-builder build`, `verify-build`; render only when the user says.91920. **Open, every session, unasked.** `scaffold` (renames old names, creates only what is missing,93 never overwrites, appends the guide once to a CLAUDE.md that lacks it, creates `style/` templates and `docs/확인할-것.md`;94 its `renamed`/`created`/`appended` in one report line), then `layout` (its warnings: a broken `analysis/` link, an95 engine path in CLAUDE.md or `style/` that is not on disk), `docs/확정사실.md`, `questions --read` and the "맞춰야 할 것" table. A stray root file (`*.conflict-*`, a scratch script) is named and moved96 to `_archive/` once the user agrees.971. **Collect.** New material → run `material` from the workspace root (items land in98 `sources/material/`). Evidence order: `sources/` → the user's analysis notes they point to →99 the user. Search before asking; ask only what only the user can answer; everything else stays100 `[확인 필요: …]`.1012. **Select.** Before any page is written, one row per item in `sources/선정.md`: 글감 · 쓰임102 (`본문` | `부록` | `제외`) · 쪽 · 이유 · 대표 (the one problem, decision or result the item leads103 with, in the material's words). Picked by signal (`references/principles.md` § Composition);104 propose the table (structure: asked). Mark each item `material log --id <id> --note105 "선정: <쓰임>" --status selected|excluded --later-step manage`. An exclusion lives only here, not106 in `확정사실.md`. `selection` reads the table back; a 쓰임 outside the three is a problem.1073. **Draft.** No skeleton yet, or a new project → write it from the `본문`/`부록` items only: one108 project as why → the user's role and decision → result; strengths chosen from the evidence.109 Each page that carries a figure or a reason cites its items on one line under its heading, `글감: <id>, <id>` (a ledger item is110 `NN_프로젝트/A1`); a reason the item does not give is written `[확인 필요: …]`, never filled in.111 The structure (blocks, order, body vs appendix) is asked; wording is not. A written112 item is marked `material log --id <id> --note "skeleton <N>쪽" --status written --later-step manage`;113 in a ledger workspace (`material ledger` lists items) `material ledger-log --item <NN_프로젝트/A1>`114 with the same flags. The ledger is never moved or rewritten.1154. **Revise, from several viewpoints.** `claims` first (each problem is an edit; `본인 확인` stays),116 then recruiter/engineer (`feedback`, `pattern`) → interviewer (`mock-interview` question-list117 mode) → the user's own bar (noah), each viewpoint's findings through `settle`. `snapshot` before118 the round's first edit (`--kind skeleton|deck|resume`); carried wording passes the decision gate (`확정사실.md`, `선정.md`);119 an order change is one `renumber`; the align table stays at the end of portfolio-skeleton.120 Rules and the order of each: `references/revise.md`.1215. **Template.** Once per template (again when it changes): `deck-builder catalog`, then122 `scaffold` copies its Palette, Type and Voice into `template/analysis.md`; add the style rules the123 pages must keep (spacing, alignment, emphasis — only what the template shows), the safe area124 (the margins nothing crosses) and the figure colours. A figure drawn with `write:diagram` takes125 its colours from `## 그림 색`.1266. **Deck.** Carry portfolio-skeleton into portfolio-deck.mdx: `sync` lists each skeleton slot missing or127 different in the deck source, `sync --write` carries them (never a slide); against `deck.catalog.md` budgets and128 `template/analysis.md`, then `deck-builder` `check` (an untouched chart slot is warned: its numbers are the template's129 sample) and `claims --file portfolio-deck.mdx` (each slide cites in its comment: `## @s3 2쪽 · 글감: <id>`). Fine style changes go through the template's run formats, never by hand in the pptx.130 `snapshot --kind deck --level … --note` only when asked (the note names its skeleton).1317. **Build, render, review.** Every round builds, then `verify-build` (each skeleton slot's text is in132 build/deck.pptx, footnote marks and picture files exist). `render` only when the user says. After a render run133 `render-review` (deck and resume alike): `keep-render`, then every page against safe area,134 readability, AI traces and the skeleton, one `검토.md` row per page. A problem seen there or in135 PowerPoint goes into `template/next-render.md` (쪽 · 칸 · 증상 · 확인 · 적용), all applied before136 the next render; fixed at once only when the user says ("고쳐", "맞춰줘"). A frame the deck cannot137 move is fixed in the template only on the user's word (`references/template-edit.md`).1388. **Feedback on the deck.** `feedback` on the rendered PDF with the material and 확정사실 in reach,139 saved to `feedback/YYYY-MM-DD.md` with its fix table. Each finding goes back to material: a claim140 the deck cannot defend → its item's `gaps.md`; a result the user restores → `result.md` labelled141 `본인 확인`. The next round starts at 2 (selection changed) or 4 (wording).1429. **Resume.** `resume-skeleton.md` (the template's section order, facts from portfolio-skeleton143 only) → `resume-deck.mdx` → `claims --file resume-deck.mdx` → `resume-builder` (catalog, build into144 the user's template, final checks, render when told, fit by `over_lines`) → `render-review` (step 7).145 The user's material never goes to git: the engine refuses a workspace git would carry.146147Always, portfolio and resume alike (user, 2026-11-10: "safe area가 있는거 시인성, 가독성이 좋은거148like-me로 ai티를 줄이는거 이건 포폴, 이력서 공통임"):149- **Safe area, readable.** Nothing crosses the margins (`layout`'s `safe_area`, from `style/레이아웃.md`, is what150 `deck-builder` uses) or touches another box; type and spacing stay the template's, one idea per line, no lone last word. Cut text, never shrink type.151- **In the user's voice.** Read `style/문체.md` and `style/표기규칙.md` before any rewrite; every sentence written into a152 skeleton or deck source goes through `write:like-me` (samples: `style/문체.md`'s own paragraphs and what the user wrote153 alone in `sources/`, never AI-edited text) to strip machine tells; then `claims`. A resume follows `style/이력서-서식.md`.154- **No traces.** Deliverables carry no AI or tool trace; the builders' final check catches it.155- **Corrections.** When the user corrects a fact, in the same turn: append to the material file156 `## 정정 — 본인 확인 YYYY-MM-DD` with their words verbatim; add one line to `확정사실.md`;157 `noah_collect` the words when noah is connected. The user's word beats the record.158- **Did it read better.** After a round's edits, score the snapshot and the new skeleton with the same rubric, then159 `compare` (`lost_ground` = worse): `references/rounds.md`. **A vague answer** ("바꾸고"): do the shared part, hold the rest, say so.160161## Output Template162163```164[결론 한 줄 — 무엇을 바꿨고 무엇이 남았나] (단계 N · skeleton vA → vB)165166바꾼 것167- N쪽 칸: "현재 문구" → "바뀐 문구" (한 줄씩, 자리만 적지 않는다)168169기본값으로 정한 것 (다르면 말씀해 주세요)170- …171172정하실 것1731. 본인만 답할 수 있는 것만174175이번 라운드 <나아짐 | 나빠짐 | 섞임 | 그대로> — 항목: 이전 → 지금 (떨어진 항목은 이전 문구와 함께)176확인한 것 (템플릿이나 원재료를 고쳤을 때: 백업 위치, check 결과)177178claims: 문제 <n> (근거 없음 n · 글감에 없음 n · 확인 필요 n · 인용 없음 n) · 본인 확인 <n> · 덱에 없는 성과 <n>179renumber/verify: <changes, warnings, problems | 안 함> · sync: <빠진 n · 다른 n> · deck check: <warnings | 안 함> · build: <함> · verify-build: <빠진 문구 n · 각주 n · 그림 파일 n> · render: <함 | 안 함>180라운드: <n>회 중 <k>회 · 종료 <stop_reason> · 게이트 gap <n> (건너뛴 단계: 이유) · 패널 <4인 | 단일> · 심사 <subagent | in-context>181기록: round-NN/ · round-check 문제 <n> · noah 판정 <값> · 열린 질문 <n> · 기준본 <stable 폴더 | 없음>182```183184Korean, conclusion first, plain headings, no bold runs, no invented names or labels; a number says what it counts.185Not built → the built files are older. A remark read one way → say the reading, ask. Held is said as held.186187## What Claude Does / What You Do188189| Claude | You |190|---|---|191| Runs material, feedback, pattern, rewrite and deck-builder as steps of one loop | Hand over material; answer only what only you know |192| Proposes the selection table (본문/부록/제외 · 대표); structure is the only wait | Approve what goes in and what each item leads with |193| Snapshots before editing, renumbers in one pass, verifies references | Approve structure changes (page count, blocks, order) |194| Runs a whole round: sync, build, verify-build; wording goes ahead and is reported | Correct facts in your own words; say when to render |195196## Related Skills197198- `material` — step 1, collects material into `sources/material/`; its status follows steps 2-3199- `feedback`, `pattern`, `rewrite`, `mock-interview` — steps 4 and 8; `write:like-me` — the voice200- `deck-builder` — steps 5-7; `resume-builder` — step 9 (docx build, PDF, fit); `render-review` — step 7201 and the resume's render; `write:diagram` — figures; `job-application-workflow` — the whole application202- `references/principles.md` — principles for any candidate and the mistakes this flow stops