Agentic Engineering

Appearance

Senior Engineer Path ​

Who it's for ​

ICs who already ship production code and want agents to own most of the typing without owning the outcome. You review PRs, write specs when it matters, and refuse to merge plausible junk.

Prerequisites ​

  • Comfortable with Git, PR review, testing, and your stack's debugger
  • Access to Claude Code (or equivalent) on a real work repo
  • Authority to add/edit CLAUDE.md and a docs/specs/ folder

Ordered modules ​

  1. Mental Models — agent as compiler of intent; quality bounded by spec
  2. Agents vs Assistants — when autonomy helps vs when it invents
  3. Why Agents Fail — catalog the failure shapes you'll see this week
  4. Prompt Anatomy — structure that survives a 200-file repo
  5. CLAUDE.md & AGENTS.md — standing orders that don't rot
  6. Specs Over Prompts — stop negotiating requirements in chat
  7. Adding a Feature Safely — the default daily playbook
  8. Reviewing AI PRs — human gates that catch agent-specific bugs
  9. Verification Strategy — structural verify, not vibes
  10. Failure Catalog — recognize thrash and shallow impl early

Exercises ​

  1. Write or prune CLAUDE.md to ≤60 lines using the CLAUDE.md template; delete anything lintable or derivable.
  2. Take one upcoming ticket; write a feature spec before any agent session.
  3. Implement that feature via Add Feature; review with the AI PR checklist.
  4. Pick one merged agent PR from the last month; list three invented requirements or drive-bys that slipped through.

Capstone ​

Ship one production feature end-to-end: filled feature spec → agent implementation → your review against the AI PR checklist → green verification → merge. File a short note: what the agent got wrong and which gate caught it (or should have).

Expected outcome ​

You default to spec-first for non-trivial work, keep repo memory files sharp, and can explain why a bad agent PR looked fine until a specific checklist item failed.

A field manual for AI-native software engineering.