Doc Co-authoring

Verified 中级 Intermediate 流程型 Process ⚡ Claude Code 专属 ⚡ Claude Code Optimized
2 min read · 90 lines

Anthropic official co-authoring workflow: context gathering → section iteration → reader testing

Doc Co-authoring

Overview

Anthropic's official collaborative document writing workflow. A three-stage structured process from context understanding through iterative refinement to reader testing.

Three-Stage Workflow

Stage 1: Context Gathering

Goal: Fully understand the requirements before writing.

  1. Meta-context questions

    • What type of document? (Technical spec / blog post / internal report...)
    • Who is the target audience? What's their background?
    • What's the expected impact of this document?
    • Format and length requirements?
  2. Info dumping

    • Ask users to provide all relevant context at once
    • Don't interrupt — let information flow naturally
    • Flag points needing later clarification
  3. Clarifying questions

    • Ask specific questions about ambiguous points
    • Close understanding gaps
    • Confirm key assumptions
  4. Exit condition: When you have sufficient understanding, move to Stage 2.

Stage 2: Refinement & Structure

Goal: Build a high-quality document section by section.

For each section, execute this loop:

Clarify → Brainstorm (5-20 options) → Curate best → Gap check → Draft → Iterate
Step Description
Clarify Confirm specific requirements for this section
Brainstorm Generate 5-20 candidate content items per paragraph/point
Curate Select the most valuable content
Gap check Verify no critical information is missing
Draft Write the initial draft
Iterate Move to next section after 3 consecutive rounds with no changes

Full document review: After all sections are complete, read through checking:

  • Does the logical flow between sections work?
  • Is terminology usage consistent?
  • Is there redundant content?

Stage 3: Reader Testing

Goal: Catch blind spots with "fresh eyes."

  1. Predict reader questions

    • What questions would a reader most likely ask?
    • List 3-5 predicted questions
  2. Sub-agent testing (if available)

    • Dispatch a fresh Claude instance to read the document
    • Collect its questions and confusion points
    • These are your blind spots
  3. Additional checks

    • Any ambiguous wording?
    • Any internal contradictions?
    • Can a non-domain-expert understand it?
  4. Iterate based on test results

    • Fix discovered issues
    • Return to Stage 2 to rewrite specific sections if needed

Tips

  • Effective guidance: Don't always agree with Claude — say "This part is wrong, change it to..."
  • Handling deviations: If writing veers off track, point directly to where to return
  • Context management: Periodically summarize current state as documents grow
  • Artifact management: Save intermediate versions in the filesystem for easy rollback

相关技能 Related Skills