Planning Foundation

.writing/ is working memory on disk, the persistent state for the current paper. Context windows are volatile and limited; the filesystem is not. Anything important gets written to .writing/ so it survives a context reset or a new session.

Directory convention

.writing/                     # gitignored, ephemeral working state
├── metadata.yaml             # authors, venue, availability, writing_profile
├── outline.md                # IMRAD structure + key claims (from outlining)
├── ideation.md               # candidate directions + verdicts (from idea, optional)
├── claims/                   # one file per section: section_<NN>_<slug>.md (claim stubs)
├── manuscript/               # LaTeX prose: <NN>_<slug>.tex (from drafting)
├── figures/                  # <fig_id>.tex/.pdf, data/, src/, briefs
├── refs.bib                  # bibliography, fills as citations resolve
├── findings.md               # discoveries, decisions, rejected alternatives
├── progress.md               # Task Status Dashboard + session log
├── agents/                   # created on demand by subagents (one dir per role)
└── archive/                  # completed papers (dir per entry + summary.md)

The outline is the paper's design spec; there is no separate design/plan document. agents/ is not created at init, each subagent makes its own subdir when dispatched.

Quick start

Before any complex writing task:

1. Create .writing/, run bash "${CLAUDEPLUGINROOT}/scripts/init-writing-dir.sh" (creates .writing/ with progress.md + findings.md seeded from templates), or create the dir and seed the two files by hand.
2. Re-read the outline before decisions, keeps the paper's goals in the attention window.
3. Update progress.md after each phase, mark status, note files written, log errors.

File purposes

Core disciplines

1. Create .writing/ first

Never start a complex writing task without it. All paper documents live here; execution status lives in progress.md's Task Status Dashboard.

2. The 2-action dispatch rule

After every ~2 read/search/explore operations, save to the right file by content type, discoveries and decisions to findings.md, status and actions and errors to progress.md. This keeps both knowledge and progress from being lost to a context reset.

3. Read before decide

Before a major decision, read the outline and relevant findings. This re-orients you when context is stale.

4. Log every error

Every error goes in progress.md (what failed, what you tried, what you observed). On repeated failure, escalate to the user rather than silently trying alternatives, the outline was designed deliberately, and self-directed workarounds bypass its intent.

Per-agent directories

When dispatching subagents, each gets .writing/agents/{role}/ with its own findings.md (discoveries) and progress.md (action log), appended across tasks, one dir per role, not per task. The orchestrator aggregates agent findings into the top-level findings.md / progress.md after each task (see scripts/aggregate-agent-findings.sh).

When to use

Use for multi-step writing, research threads, and anything spanning many tool calls or sessions. Skip for simple questions, single-file edits, and quick lookups.

The 5-question reboot test

If you can answer these from .writing/, your state management is solid: Where am I? (Task Dashboard) · Where am I going? (remaining phases) · What's the goal? (outline) · What have I learned? (findings.md) · What have I done? (progress.md).

Scripts

At the plugin root, invoked as bash "${CLAUDEPLUGINROOT}/scripts/<name>":

• init-writing-dir.sh, initialize .writing/ with findings.md + progress.md
• check-writing-state.sh, report state: missing | empty | active | complete
• writing-reset.sh, reset active state, preserve archive/
• snapshot-save.sh, copy active files to a target dir (used by archiving)
• aggregate-agent-findings.sh, merge subagent findings into the top-level files
• detect-base-branch.sh, detect-test-command.sh, unique-filename.sh, project detection helpers

Anti-patterns