Write Docs

Write Docs

Turn session knowledge into structured, readable technical documentation. You are the controller — you plan the document, delegate sections to focused subagents, assemble the result, and polish it.

When This Triggers

You've been working through something — debugging, setting up infrastructure, exploring an API, configuring a tool — and now the user wants a written artifact that captures what was learned. The documentation lives beyond this session: other people (or future-you) will read it cold.

The Process

digraph write_docs {
    rankdir=TB;
    "Gather context from session" [shape=box];
    "Clarify scope with user" [shape=diamond];
    "Propose structure (TOC)" [shape=box];
    "User approves structure?" [shape=diamond];
    "Dispatch section writers" [shape=box];
    "Assemble draft" [shape=box];
    "Self-review pass" [shape=box];
    "Humanizer pass" [shape=box];
    "Present to user" [shape=box];
    "User feedback?" [shape=diamond];
    "Revise" [shape=box];
    "Write final file" [shape=box];

    "Gather context from session" -> "Clarify scope with user";
    "Clarify scope with user" -> "Propose structure (TOC)";
    "Propose structure (TOC)" -> "User approves structure?";
    "User approves structure?" -> "Propose structure (TOC)" [label="revise"];
    "User approves structure?" -> "Dispatch section writers" [label="approved"];
    "Dispatch section writers" -> "Assemble draft";
    "Assemble draft" -> "Self-review pass";
    "Self-review pass" -> "Humanizer pass";
    "Humanizer pass" -> "Present to user";
    "Present to user" -> "User feedback?";
    "User feedback?" -> "Revise" [label="changes needed"];
    "Revise" -> "Present to user";
    "User feedback?" -> "Write final file" [label="approved"];
}

Step 1: Gather and Clarify

Before proposing structure, understand what you're documenting:

Extract from session context:

• What was built, configured, or discovered
• Key decisions and their rationale
• Commands that worked (and gotchas that didn't)
• Dependencies, prerequisites, environment details

Ask the user (one at a time, skip what's obvious from context):

• Who reads this? (team members, open source users, future self)
• What do they need to do after reading? (set up, understand, decide, troubleshoot)
• Where does this live? (README section, standalone guide, wiki page, repo docs/)
• What format? (tutorial with steps, reference doc, explanation, how-to guide)

These four documentation types come from Divio's documentation system — pick the right one and the structure follows naturally:

• Tutorial: learning-oriented, follows a path, "do this then this"
• How-to guide: task-oriented, solves a specific problem, assumes competence
• Explanation: understanding-oriented, gives context and background
• Reference: information-oriented, accurate and complete, like a dictionary

Step 2: Propose Structure

Present a table of contents with one-line descriptions of what each section covers. Keep it lean — documentation that's too long doesn't get read.

Format your proposal like:

## Proposed structure

**File:** `docs/setup-guide.md`
**Type:** How-to guide
**Audience:** Team developers setting up locally

1. **Prerequisites** — what you need installed before starting
2. **Installation** — clone, install deps, configure env
3. **Configuration** — the three env vars and what they control
4. **Verification** — how to confirm it's working
5. **Troubleshooting** — the two gotchas we hit during setup

Estimated length: ~200 lines. Sound right?

Wait for approval. Adjust if the user wants sections added, removed, or reordered.

Step 3: Delegate Sections

Dispatch one subagent per section (or group small related sections). Each subagent gets:

1. The section brief — what to cover, approximate length, level of detail
2. Audience context — who reads this, what they already know
3. Source material — specific facts, commands, code snippets, decisions from the session
4. Style guidance — see the Style section below

Subagent prompt template:

Write a documentation section. Return ONLY the markdown content (no fences wrapping it).

**Section:** [title]
**Brief:** [what to cover]
**Audience:** [who, what they know]
**Type:** [tutorial / how-to / explanation / reference]
**Source material:**
[paste the relevant facts, commands, code, decisions]

**Style rules:**
- Write for someone reading this cold — no "as we discussed" or "as mentioned above"
- Use second person ("you") for instructions
- Code blocks with language tags
- One idea per paragraph
- Short paragraphs (3-4 lines max)
- Prefer concrete examples over abstract descriptions
- If a step can fail, say what failure looks like and what to do

Parallelization: Independent sections can be dispatched in parallel. Sections that reference each other (e.g., "Configuration" needs to know what "Installation" set up) should be sequential — or give the later section the content from the earlier one.

Step 4: Assemble and Review

Combine section outputs into one document. Then do two passes:

Self-review (do this yourself, inline):

• Consistent heading levels and terminology throughout
• No contradictions between sections
• No gaps where "the reader needs to know X but no section covers it"
• Logical reading order — does each section build on the previous?
• Code examples actually match the described steps
• No orphan references ("see below" pointing nowhere)

Humanizer pass: Read the assembled draft through the lens of the humanizer skill. Specifically check for:

• Significance inflation ("crucial", "vital", "key")
• Promotional language ("powerful", "seamless", "robust")
• Copula avoidance ("serves as" instead of "is")
• Filler phrases ("in order to" → "to")
• Sycophantic intros ("Great question!")
• Em dash overuse
• Bolded list headers where flowing prose would work better

Fix issues inline. The goal is documentation that reads like a competent engineer wrote it, not an AI.

Step 5: Present and Iterate

Show the full draft to the user. Don't ask "should I continue?" — present it and wait.

If the user has feedback:

• For small fixes: apply directly
• For section rewrites: re-dispatch that section's subagent with the feedback as additional guidance
• After changes, do another quick humanizer check on the modified sections

When approved, write the final file to the agreed path.

Style Defaults

Unless the user specifies otherwise:

• Headings: sentence case ("Getting started" not "Getting Started")
• Code blocks: always tagged with language (bash, json, typescript, etc.)
• Commands: show the command AND its expected output when useful
• Links: inline [text](url) not reference-style
• Lists: use when items are genuinely parallel; don't force prose into bullets
• Length: as short as possible while remaining complete — every sentence earns its place
• Tone: direct, confident, no hedging ("this will" not "this should probably")
• Structure: front-load the useful information — don't make readers wade through context to find the command they need

Choosing Not to Subagent

For short documents (under ~100 lines, 2-3 sections), skip the subagent dispatch — write it directly. The subagent pattern pays off when:

• The document has 4+ substantial sections
• Different sections need different expertise or source material
• The total length would be 150+ lines
• Writing it all inline would bloat your context

Output

The final artifact is always a .md file written to disk. Confirm the path with the user before writing. Common locations:

• docs/ in the project root
• README.md (appended or replaced section)
• Wiki or standalone file

After writing, show the file path and a one-line summary of what was documented.