Automate everything you do
If you do something more than twice, an agent should do it. Automation only pays off when it has an owner and fails loudly. A silently broken release-notes generator is worse than a human writing them.
Pick targets by frequency times pain, not by what is fun to automate:
| Artifact | How agents handle it | Review needed |
|---|---|---|
| Code | Claude Code, Cursor, Copilot | Always |
| Documentation | Generated from code, types, and tests | Light, spot-check |
| Diagrams | Mermaid, PlantUML, D2 from the actual code | Check against reality |
| API specs | OpenAPI from code, or code from OpenAPI | Full, other teams consume it |
| Database migrations | Scripts from schema changes | Full, irreversible |
| Commit messages, PR descriptions | /commit, /pr skills | Light |
| Release notes | From commits and merged PRs | Light, customers read them |
| Runbooks | From incident history and infra code | Full, used under pressure |
| Test data | Synthetic data with Faker plus AI | Check for real PII |
| Code reviews | Claude GitHub app on every PR | Humans still own approval |
| Refactoring | Agent-proposed, human-approved | Full |
| Translations, alt text, ARIA | Generated per change | Spot-check |
| Slides, meeting notes, email drafts | Claude, NotebookLM, Gamma | Your name is on it |
The shift in practice:
Before: "I need to write documentation for this API"
After: "Generate the OpenAPI spec from src/routes/, then markdown docs from the spec"
Before: "I need an architecture diagram"
After: "Read the infrastructure code and generate a Mermaid diagram of it"
Before: "I need slides for the team meeting"
After: "Summarize this week's merged PRs into a 5-slide update"Example prompts:
Documentation:
"Document every endpoint in src/routes/: request and response examples, error codes,
auth requirements. Output markdown."
Release notes:
"Release notes for v2.4.0 from commits since v2.3.0. Sections: Features, Fixes,
Breaking Changes."
Runbooks:
"Write the database failover runbook: commands, expected output, escalation contacts.
Optimize for scanning during an incident."Tools for non-code artifacts:
| Category | Tools |
|---|---|
| Diagrams | Mermaid, PlantUML, D2, Excalidraw |
| Presentations | Gamma, Claude plus reveal.js |
| Images | Gemini, OpenAI image models, Midjourney, Flux |
| Video and audio | Runway, HeyGen, ElevenLabs, Descript |
| Docs | Notion AI, Mintlify |
| Synthetic data | Faker plus an LLM, Gretel, Mostly.ai |
Staff rules for automation:
- Every automation has an owner and a failure signal
- Artifacts other teams consume (specs, migrations, release notes) get code-level review
- Consolidate on a few vetted tools with known data-handling terms. Don't add a new SaaS for each task
- Never automate a process nobody has run by hand. If you do, you scale its mistakes