Paper Writing Memory Manager

Maintain the paper's dynamic writing memory. This skill is the global state layer for nonlinear paper writing: it records what is being written, why wording changed, which claims appear where, which sections are stale, and what must be updated when evidence, claims, figures, tables, captions, or style decisions change.

Use this skill for:

• initializing paper/.agent/writing-memory/
• recovering the current writing state at the start of a session
• recording section status, paragraph status, caption/table/figure edits, and open writing threads
• mapping claims, evidence, results, figures, tables, captions, and paper locations into a dependency map
• deciding the impact of a local edit or new experiment result across the whole draft
• marking affected sections stale after claim/evidence/result changes
• recording style, terminology, rhythm, and venue-specific writing decisions
• handing off work between section-specific writing skills

Do not use this skill to write the prose itself. Use paper-writing-assistant and section-specific writing skills for drafting. Do not use it as the claim/evidence source of truth; use paper-evidence-board for claim/evidence status. Do not use it for long-term cross-project memory; use research-project-memory for project-level state.

Skill Directory Layout

<installed-skill-dir>/
├── SKILL.md
├── references/
│   ├── memory-schema.md
│   ├── update-protocol.md
│   ├── writing-layers.md
│   └── impact-propagation.md
└── templates/
    ├── writing-state.md
    ├── section-ledger.md
    ├── dependency-map.md
    ├── edit-impact-log.md
    ├── writing-style.md
    ├── style-and-terminology.md
    ├── open-writing-threads.md
    └── session-notes.md

Progressive Loading

• Always read references/memory-schema.md and references/update-protocol.md.
• Read references/writing-layers.md before nontrivial prose edits, style/terminology changes, layout-motivated rewrites, venue adaptation, or final polish.
• Read references/impact-propagation.md when a claim, result, table, figure, caption, title, abstract, or limitation changes.
• Use templates when initializing missing files under paper/.agent/writing-memory/.
• Use templates/writing-style.md when initializing paper/.agent/writing-style.md.
• Read paper/.agent/writing-contract.md, paper/.agent/writing-style.md, paper/.agent/writing-style-lessons.md, paper/.agent/notation-contract.md, paper/.agent/paper-evidence-board.md, paper/.agent/evidence-completion-plan.md, paper/.agent/result-asset-provenance.md, paper/.agent/provisional-results.md, and paper/.agent/consistency-report.md when present.
• Read root memory/claim-board.md, memory/evidence-board.md, memory/provenance-board.md, memory/handoff-board.md, and memory/phase-dashboard.md when present and the writing change affects project-level state.
• Read target draft files only as needed for the requested section, paragraph, caption, table, or figure.

Core Principles

• Writing is nonlinear; memory must preserve active threads, stale areas, and dependencies across sessions.
• A local edit can have global consequences. Record semantic impact, not just textual diff.
• The dependency map should make cross-section consistency tractable.
• Writing memory is agent-private. If the active paper branch is Overleaf/coauthor-visible, anonymous-submission, arXiv/public, or camera-ready public source, ensure .agent/ is ignored/untracked or use an agent-private paper worktree/root memory instead.
• Section status should distinguish stable, draft, stale, blocked, missing, and needs-review.
• Style and terminology decisions are part of writing memory because they affect future edits.
• Writing edits should declare their active layer: layout, surface-fluency, argument, technical-consistency, style-consistency, venue-adaptation, or final-polish.
• Writing style should evolve as memory: record one-off lessons, promote repeated preferences into project contracts, and only then generalize them into reusable skill rules.
• Do not duplicate full paper text or long experiment reports. Store pointers, locations, IDs, and concise decisions.
• Every substantial writing action should leave a small memory trace: what changed, why, what it affects, and what remains open.
• If a writing change affects project-level claim lifecycle, evidence provenance, phase gates, or cross-module work, summarize that impact in root memory/ through research-project-memory rather than hiding it in paper-local memory.

Step 1 - Locate or Initialize Writing Memory

Find the paper root:

• paper/ under a project control root
• current directory if it is the paper repo
• user-provided path

The memory directory is:

paper/.agent/writing-memory/

If there is no paper/ directory and the current directory is the paper repo, use:

.agent/writing-memory/

Initialize missing files from templates/:

• writing-state.md
• section-ledger.md
• dependency-map.md
• edit-impact-log.md
• style-and-terminology.md
• writing-style.md when a paper-wide style contract is useful
• open-writing-threads.md
• session-notes.md

Step 2 - Build or Refresh the Writing Snapshot

Read references/memory-schema.md.

Create a compact snapshot:

## Writing Memory Snapshot
- Venue:
- Paper archetype:
- Current positioning:
- Active writing focus:
- Stable sections:
- Draft sections:
- Stale sections:
- Blocked sections:
- Active claims:
- Recent evidence/result changes:
- Open writing threads:
- Highest-risk dependencies:
- Recommended next writing action:

Use local paper artifacts first. If no writing contract or evidence board exists, record that as a blocker and route to the appropriate skill.

Step 3 - Update Section and Dependency Memory

For any writing action, update:

• section-ledger.md: status, claim roles, evidence dependencies, figures/tables, stale paragraphs, next action
• dependency-map.md: claim/result/evidence/asset IDs to paper locations
• style-and-terminology.md: new or changed terms, method names, caption style, claim-strength rules
• paper/.agent/writing-style.md or writing-memory/style-and-terminology.md: active writing layer decisions, style lessons, user preferences, protected invariants, and project contracts
• open-writing-threads.md: unresolved writing questions and blockers
• edit-impact-log.md: semantic change and affected locations
• session-notes.md: what happened in the current session

Keep entries concise and link to files/sections instead of copying long prose.

Step 4 - Propagate Impact

Read references/impact-propagation.md when something changes.

Classify the changed object:

• claim
• result/evidence
• table
• figure
• caption
• method notation
• title/abstract/contribution
• related-work boundary
• limitation/scope
• terminology/style

Then mark affected locations:

• title
• abstract
• contribution bullets
• introduction
• method
• experiments/results
• figures/tables/captions
• related work
• limitations
• conclusion
• appendix

For each affected location, mark status: update-needed, verify, stable, blocked, or cut-candidate.

Step 5 - Route the Next Writing Skill

Route by active thread:

• title/abstract/contribution -> abstract-title-contribution-writer
• introduction -> paper-introduction-argument-writer
• method explanation -> method-section-explainer
• results prose -> experiment-story-writer
• related work -> related-work-positioning-writer
• limitations/scope -> limitations-scope-writer
• result gap -> paper-evidence-gap-miner
• CSV-derived table/figure -> paper-result-asset-builder
• broad prose edit -> paper-writing-assistant
• full consistency pass -> paper-draft-consistency-editor

Step 6 - Close the Session

At the end of a writing session, update session-notes.md and writing-state.md:

• what changed
• why it changed
• which dependencies were added or invalidated
• which sections are now stale or stable
• which evidence gaps remain
• which project-level provenance, handoff, claim lifecycle, or phase-dashboard updates are needed
• next recommended action

If project memory exists, create or update memory/handoff-board.md when a writing session produces work for another module, such as a missing-evidence request for paper-evidence-gap-miner, an asset request for paper-result-asset-builder, a new-compute request for experiment-design-planner, or a consistency request for paper-draft-consistency-editor.

Final Sanity Check

Before finalizing:

• writing memory files exist and are compact
• active claims map to paper locations or are marked unplaced
• section statuses reflect current draft reality
• stale locations are explicitly marked after evidence or claim changes
• style and terminology decisions are recorded
• nontrivial edits have an active writing layer and protected invariants recorded
• open writing threads have next actions
• project-level handoffs are created when writing needs evidence, assets, diagnosis, new compute, or consistency review
• other skills know what to do next