ethos
← Back to Projects
📚

Docs

Keep internal docs in sync with code

Thesis

Internal docs are outdated or missing. AI can watch for changes and flag when docs need updates.

The Problem

Documentation is always out of date: Someone writes a README. The code changes. The README doesn't. Six months later, new developer follows the docs and everything breaks. "Oh yeah, we don't do it that way anymore." Nobody has time to update docs. So they rot. And onboarding suffers, bugs happen, knowledge lives only in people's heads.

Implementation Approaches

Approach 1

Git Watcher + Flag System (Recommended)

Recommended

Monitor code changes and flag docs that might be stale

⏱️2 weeks
📊Medium complexity

Implementation

  • GitHub/GitLab integration watches commits
  • Map docs to code sections they describe
  • AI detects: did this change affect related docs?
  • Flag docs for review with context
  • Suggest specific updates when possible

Pros

  • +Proactive, catches staleness as it happens
  • +Works with existing doc structure
  • +Clear action items for doc updates
  • +Flywheel: learns what changes matter for docs

Cons

  • Mapping docs to code is imperfect
  • May generate false positives
  • Requires docs to exist in the first place
Approach 2

Doc Generator from Code

Auto-generate documentation from codebase

⏱️2 weeks
📊High complexity

Implementation

  • Analyze codebase structure and patterns
  • Generate: architecture overview, API docs, setup guides
  • Keep in sync via regeneration on changes
  • Diff-based updates to preserve manual additions

Pros

  • +Docs always match code by definition
  • +Reduces manual documentation burden
  • +Consistent format across projects
  • +Can generate from scratch for undocumented code

Cons

  • Generated docs often lack context/reasoning
  • Quality depends on code quality
  • May overwrite valuable manual content
Approach 3

Knowledge Q&A Layer

Chatbot that answers code questions from docs + code

⏱️1 week
📊Low complexity

Implementation

  • Index docs and codebase
  • Developers ask questions in natural language
  • AI answers with sources
  • Track: what questions aren't covered?

Pros

  • +Works even with incomplete docs
  • +Surfaces what's actually needed
  • +Lower bar than full doc maintenance
  • +Shows gaps via unanswered questions

Cons

  • Doesn't fix underlying doc problem
  • RAG quality can be inconsistent
  • Developers might not ask

Validation Plan

Hypothesis to Test

Dev teams will pay $29/mo to keep docs in sync with code automatically

Validation Phases

1

Staleness Audit

1 week
  • Audit docs in 3 active repos
  • Compare doc content to current code
  • Find: what's stale, what's missing
  • Estimate: how much time to fix manually
Find significant staleness in 3/3 repos
2

Watcher MVP

2 weeks
  • Build GitHub integration
  • Implement doc-to-code mapping
  • Test staleness detection on real repos
  • Measure: true positive rate
Catch 80%+ of doc-affecting changes with < 20% false positives
3

Team Beta

2 weeks
  • Deploy to 3 dev teams
  • Integrate with their doc workflow (Notion, Confluence)
  • Track: flags generated, flags acted on
  • Validate $29/mo pricing
Teams act on 50%+ of flags, want to continue

Kill Criteria

Stop and move on if any of these become true:

  • Too many false positives make it noisy
  • Doc-to-code mapping is too unreliable
  • Teams ignore flags, don't value doc currency
  • Competitive tools already do this well