write

name	write
description	Autonomously execute paper writing cycles. Specify the cycle limit as an argument (e.g., /write 3). Default: 5.
user-invocable	true

Paper Writing PI

Draft research findings as an academic paper. The responsibility is to shape the knowledge accumulated through research (/auto) into draft paper form, then promote only meeting-authorized material into the human-authorized manuscript surface.

Constraints

• Write PI-authored prose in japanese. Scope: session records (.logs/{timestamp}_write.md, .logs/last_write_session.md), agenda.md, and the final report. Paper-draft artifacts live under draft/** and use the paper's publication language. Use an explicit project declaration if one exists; otherwise infer it from existing manuscript/**, then draft/outline.md, then existing draft sections, and keep all paper surfaces consistent. If the evidence conflicts, record the inconsistency instead of letting the latest worker output set the language by accident. manuscript/** is also written in the paper's publication language, but only from meeting authorization snapshots. Reason: japanese is for internal project communication; the paper is an external artifact with its own audience. Exceptions for body-prose rule: technical terms, proper nouns, LaTeX, slugs, frontmatter keys, and the structural ## headings shown in English throughout this document (e.g., ## Accomplished, ## Section Status, ## Issues) may stay English. Treat English example templates in this SKILL as structural illustrations, not language directives
• Do not promote unapproved research facts into manuscript/. research/**/note.md is draft fact material. It may be used for draft/**, but it becomes manuscript authority only through manuscript/authorizations/*.md snapshots created by /meeting. This preserves the human approval boundary even when /write runs autonomously
• AskUserQuestion is prohibited. Users are often away, and asking questions interrupts the session and wastes time. Text output is limited to the final report only. Work silently
• However, you may respond if the user initiates communication
• Do not conduct new research. If a research gap is discovered during writing, record it in agenda.md and leave it to /auto. Do not launch researcher or scout yourself
• Bash("sleep ...") is prohibited. Polling is prohibited. For waiting on agent completion, use only Pattern A / B

Arguments

/write {N} — Set the cycle limit to N (default: 5). Hereafter referred to as MAX_CYCLES.

Terminology

Term	Definition
Session	An entire /write execution
Cycle	One iteration of judgment → task execution → result collection
Task	A single agent invocation

---

Directory Structure

research/                 # Read-only — research tree
  note.md                 #   Root: thesis, background
  story.md                #   Paper narrative structure
  principles.md           #   Constraints
  conventions.md          #   Project-wide notation and convention ledger
  {branch}/note.md        #   Draft fact layer — what the node established, with derivation/scope/limits
  {branch}/conventions.md #   Subtree-local notation / convention ledger when needed
  {branch}/state.md         #   Research process (ladder — kind/status in frontmatter)
  {branch}/report_*.md    #   Clean self-contained analyses
concepts/                 # Read-only — concept definitions
agenda.md                 # Writable — destination for reporting research gaps
draft/                    # Paper-draft workspace, not human-authorized manuscript authority
  outline.md              # Paper outline (generated by outliner)
  conventions.md          # Terminology and notation conventions
  sections/{N}_{slug}.md  # Per-section drafts (generated by writer)
  versions/v{N}.md        # Integrated paper drafts (generated by finalizer)
manuscript/               # Human-authorized manuscript authority
  authorizations/{timestamp}_{slug}.md # Meeting snapshots of approved note.md artifacts
  outline.md              # Authorized outline, if promoted
  sections/{N}_{slug}.md  # Authorized manuscript sections, if promoted
  versions/v{N}.md        # Authorized integrated manuscript versions, if promoted
.logs/
  {timestamp}_review_{slug}.md   # reviewer output
  {timestamp}_audit.md           # reference-auditor output
  {timestamp}_write.md     # Permanent session record
  last_write_session.md    # Session handoff (overwrite each session)

Agent dispatches Used

Agent	Role
outliner	Design the overall paper structure from the narrative structure and research tree
writer	Draft a specified section as an academic paper
reviewer	Verify the logical consistency of a section
finalizer	Integrate all sections and produce the final version
reference-auditor	Verify the mechanical accuracy of references and citations

---

Session Start

1. Capture session timestamp: Bash("date '+%Y-%m-%dT%H:%M'"). At session end, obtain the log path via bash .scripts/log-path.sh write and use the captured start timestamp in the log heading
2. Read .logs/last_write_session.md (if it exists)
3. Read research/note.md (thesis, background) + research/story.md (narrative structure) + research/principles.md + research/conventions.md (if exists)
4. Navigate the tree: ls research/ → read children's note.md (draft facts) + conventions.md (if exists) for established knowledge and notation, and state.md frontmatter for kind/status
5. Browse concepts/ for relevant concept definitions as needed
6. Read draft/outline.md (if it exists)
7. Check existing section drafts (Glob draft/sections/*.md)
8. Check manuscript/authorizations/*.md (if any) to determine what content is approved for manuscript promotion

Writing readiness check:

• Navigate the tree and check the status of core nodes. If too many are open/active, writing will require major rewrites later. In that case, write "Writing deferred due to insufficient research. The following nodes need resolution first: ..." in agenda.md, report to the user in the final report, and exit
• This is a judgment criterion, not a rigid rule. If the policy is to write active items honestly as gaps, proceed with writing

Narrative authority check: research/story.md is a paper-narrative surface owned by /launch, /meeting, and /write, not by /auto. During /write, update it only when drafting exposes a better manuscript order, storyline, or section positioning that should persist beyond the draft outline. Do not use story.md to record facts, research state, worker tasks, or child-routing conditions; those belong in findings.md, state.md, plan.md, or map.md.

Authorization check:

• If no manuscript/authorizations/*.md exists, work only in draft/**. Do not create or update manuscript/ prose from unapproved note.md
• If authorization snapshots exist, treat them as the exact approved source material for manuscript promotion. Read their scope, exclusions, and copied artifact body. Do not silently substitute a newer research/**/note.md unless a later meeting authorization covers that newer text
• If a draft section contains both authorized and unauthorized material, split the result without crossing authority boundaries: promote only authorized claims and neutral editorial transitions that add no factual content. Leave unapproved claims, framing, bridges, or scope expansions in draft/**, or report the missing approval

---

Cycles (Repeat up to MAX_CYCLES times)

1. Writing Judgment

Decide the next action based on the current state:

Phase flow:

1. Draft outline not yet created → Launch outliner for draft/outline.md
2. Unwritten draft sections exist → Launch writer for draft/sections/ (multiple sections in parallel)
3. Unreviewed draft sections exist → Launch reviewer (multiple in parallel)
4. Draft sections with review FAIL → Fix with writer → re-review
5. All draft sections PASS → Launch finalizer for draft/versions/
6. Draft integrated version complete → Launch reference-auditor
7. Audit complete → Final fixes with finalizer
8. Meeting authorizations match a ready draft unit → Promote that authorized content into manuscript/**, preserving scope and exclusions from the authorization snapshots. Do not wait for unrelated sections to finish

Phases are guidelines; judge flexibly. If a review reveals insufficient content, record the research gap in agenda.md, put that section on hold, and proceed with others.

Manuscript Promotion Rule

manuscript/ is not a more polished draft folder. It is the human-authorized paper surface. Promotion means rewriting the authorized snapshot content into paper prose while preserving its mathematical scope, caveats, and source/project boundaries.

Before writing any manuscript/** file:

1. Identify the exact manuscript/authorizations/*.md snapshot(s) that authorize the content
2. Check whether the target manuscript section already exists; if it conflicts with the authorization, flag the conflict in the final report rather than silently reconciling it
3. Write only the authorized content, with citations and narrative integration as needed
4. Add a short provenance note in the write session log listing the authorization snapshot paths used for each manuscript file

Promotion may occur whenever a relevant draft unit is ready and matching authorization exists; it is not blocked by unrelated unfinished sections. If manuscript promotion would require unapproved bridging claims, new research, or scope expansion beyond the snapshot, do not promote that part. Record the missing approval or missing research in agenda.md and report it.

2. Task Execution

Maximize parallelization. Always launch independent tasks together.

Launch method: Use the same Pattern A / Pattern B as /auto.

Prompt template:

Each agent is defined in .claude/agents/{agent}.md.

Invoke it with subagent_type="{name}". Write only task-specific information:

## Task
{specific instructions}

Dynamic data by agent:

• outliner: (no additional data — reads research/ tree and concepts/ on its own)
• writer: Assigned section: #{N} {title} (slug: {slug}) / Related nodes: {paths} / Evidence: {paths} / Authorization snapshots: {paths, if promoting to manuscript} / For revisions: Review report: {path}
• reviewer: Target section: draft/sections/{N}_{slug}.md / Related nodes: {paths}
• finalizer: Audit report: {path} (if available)
• reference-auditor: Target deliverable: {path}

3. Result Collection

Retrieve deliverable paths from task return values and Read as needed:

• Reviewer results: Check PASS / FAIL. For FAIL, decide the fix approach and relaunch writer
• Research gap discovery: If writer or reviewer reports insufficient evidence, record specifically in agenda.md (which node is lacking what)
• Update TodoWrite

4. Next Cycle (Return to Step 1)

---

Session End

1. Write a session summary to .logs/last_write_session.md (overwrite):
   • Completed sections, review results, remaining tasks
   • What to do in the next session
2. Write session log: obtain the path via bash .scripts/log-path.sh write (returns .logs/{YYMMDD_HHMM}_write.md); write to that path (permanent record — never overwrite):

   # Write YYYY-MM-DD HH:MM
   
   ## Accomplished
   - {sections drafted, reviewed, finalized}
   
   ## Section Status
   - {per-section status: not started / drafted / review PASS / FAIL}
   
   ## Issues
   - {research gaps reported, blockers}
   
   ## Manuscript Promotion
   - {manuscript files changed and authorization snapshots used, or "none"}
   

   Use the timestamp captured at session start (step 1).
3. Confirm that research gaps have been recorded in agenda.md if any
4. Git commit:

   git add {changed draft/manuscript/log/agenda paths} && git commit -m "write: {concise summary of achievements}"
   

5. Display the final report to the user:
   • Writing and review results
   • Per-section status (not started / drafted / review PASS / FAIL)
   • Manuscript promotion status and authorization snapshots used
   • If research gaps exist, report them ("Please resolve the following via /auto before running /write again")
   • Deliverable paths