doc

Documentation Skill

Config-driven documentation scaffolding. Config lives inside the skill at config/defaults.yaml — no repo-root files needed.

Commands

Command	What to run
start <name>	bash ${AGENTS_SKILLS_DIR}/scripts/start.sh "<name>"
expt <name>	bash ${AGENTS_SKILLS_DIR}/scripts/expt.sh "<name>"
plan <index> <title>	bash ${AGENTS_SKILLS_DIR}/scripts/plan.sh <index> "<title>"
finding <index> <title>	bash ${AGENTS_SKILLS_DIR}/scripts/finding.sh <index> "<title>"
ckpt <index> <description>	bash ${AGENTS_SKILLS_DIR}/scripts/ckpt.sh <index> "<description>"
research <index> <topic>	bash ${AGENTS_SKILLS_DIR}/scripts/research.sh <index> "<topic>"
review <index> <title>	bash ${AGENTS_SKILLS_DIR}/scripts/review.sh <index> "<title>"
learn <index> <domain> <title>	bash ${AGENTS_SKILLS_DIR}/scripts/learn.sh <index> "<domain>" "<title>"
list	bash ${AGENTS_SKILLS_DIR}/scripts/list.sh
status <index>	bash ${AGENTS_SKILLS_DIR}/scripts/status.sh <index>
resume <index>	bash ${AGENTS_SKILLS_DIR}/scripts/resume.sh <index>

start = scaffold docs/ + create experiment. Use this when beginning new work. expt = create experiment only. Use when docs/ already exists.

Agent Behavior — READ THIS

You are responsible for tracking the active experiment per conversation. When the user creates an experiment or references one, remember its index. Pass it to every script call — scripts are stateless and parallel-safe.

When to use each command

User says...	You do...
"Let's investigate X" / starts new work	/doc start "<name>" — remember the returned index
"Plan this" / "How should we approach"	If no recent plan covers this work: /doc plan <idx> "<title>" then write plan content. If updating/refining a recent plan: edit that same plan file.
User enters plan mode	STOP. Do NOT use .claude/plan.md. Use the existing relevant plan when this is an update; otherwise run /doc plan <idx> "<title>" and write there. Plans live in experiment folders, never in ephemeral storage.
You discover something significant	/doc finding <idx> "<title>" then write finding content
Conversation wrapping up / context switch	/doc ckpt <idx> "<short-label>" — short label only (2-5 words), then write Complete Agent Handoff content with a Mermaid diagram into the file. Do this automatically without being asked.
"Research X" / "Look into Y"	/doc research <idx> "<topic>" — write prompt, do research, write results
"Review X" / "Check this output"	/doc review <idx> "<title>" then write review content with verdict
"Remember this" / "Key learning"	/doc learn <idx> <domain> "<title>" — write distilled insight
"Continue experiment 19" / "Pick up"	/doc resume <idx> — reads meta + latest plan + Last 5 checkpoints. Set as active for this conversation.
"What experiments exist?"	/doc list
"What's the state of experiment 19?"	/doc status <idx>

Context resolution (when user doesn't give an index)

1. Experiment created or resumed in this conversation → use it
2. User referenced one by name/number → use it
3. User's request clearly relates to one experiment → use it
4. Ambiguous → ask: "Which experiment?" then run /doc list to show options

You resolve context. Scripts just take a number.

Auto-behaviors

• Checkpoint automatically at natural stopping points — end of conversation, before context switch, after a milestone. Don't wait to be asked.
• Plans go in experiments. NEVER write plans to .claude/plan.md or any ephemeral storage.
• Update the same recent file when the user is revising, refining, or extending that artifact. Recent active doc = plan/finding/review/research file you created or resumed in this conversation, or the latest relevant file from /doc resume. Edit it directly.
• Create a new numbered file when the topic, decision point, milestone, or artifact purpose changed. In short: create a new numbered file for distinct work. Examples: new implementation approach after old plan is obsolete; separate finding; new review; new research question; checkpoint snapshots.
• Agent takes the call. If unclear whether this is update vs new: prefer editing the recent active doc when it still represents current work; create new only when history should preserve a distinct version.
• Resume before continuing. When picking up an existing experiment, run /doc resume first to load context, including Last 5 checkpoints for continuity.

Checkpoint content — MANDATORY

The ckpt command arg is a short label (2-5 words), NOT the checkpoint content. After the script creates the file, you MUST write the body into it using this structure:

```mermaid
graph TD
    A[Starting state] --> B[What was done]
    B --> C[Current state]
    C --> D[Next steps]
```

## What
What was accomplished. Bullet points, brief.

## Key Takeaways
Surprising findings, confirmed/rejected hypotheses, numbers that matter.

## Current State
Exact repo state: branch/commit if relevant, changed files, active experiment index, active plan/review/research files, commands already run, test results.

## Issues
Blockers hit, workarounds used, unresolved problems. Include exact error messages if relevant.

## Decisions
Key choices made and why. What was considered and rejected.

## Inferred Priority
Your inferred next priority based on conversation and repo state. Mark as "inferred" unless user stated it explicitly.

## Next
Concrete next steps with enough detail that a fresh agent can continue without reading chat history.
Include file paths, command examples, what to watch out for, and first command to run next.

The checkpoint is a Complete Agent Handoff. It must contain a Mermaid diagram plus enough detail that a fresh agent can continue without reading chat history or asking what happened. Be brief but complete — file paths, key numbers, decisions, current state, priority, what worked, what didn't, and exact next command.

DO NOT cram content into the description argument. Keep the arg short: /doc ckpt 7 "v1 analysis complete" then edit the file.

Experiment Structure

experiments/NNN-{name}/
├── .meta.json          auto-updated by scripts
├── plans/              multiple plans, numbered
├── findings/           multiple findings, numbered  
├── checkpoints/        progress snapshots, numbered
├── review/             structured reviews, numbered
└── research/           prompt + response pairs, numbered

Setup

/doc start my-investigation — that's it. Scaffolds docs/ and creates the first experiment.

To customize folder structure or naming: edit config/defaults.yaml inside the skill.

Rules

1. Scripts only create empty templates (frontmatter + headings). YOU must write the actual content into the file after the script runs. Never treat the script call as the finished step.
2. Always use the scripts for file creation. Never manually create experiment dirs or edit .meta.json.
3. Index is just the number. 23 not 023 not 023-docs-and-structs.
4. No auto-commit. Scripts scaffold files — you or the user decide when to commit.
5. Plans go in experiments. Never in .claude/plan.md. This is non-negotiable.
6. Recent active docs are editable; historical docs are append-only. Update the same recent file for revisions to current work. Create a new numbered file for distinct topics, milestones, decisions, or snapshots.
7. Valid learnings domains are listed in config/defaults.yaml. learn will error on invalid domains.
8. Keep titles short. Titles become filenames. Use 2-5 words max — concise and descriptive. "auth-flow" not "investigation-into-authentication-flow-reliability-issues". The numbering prefix handles ordering; the title just needs to identify the topic at a glance.

$ARGUMENTS