• The AI Agent Operator’s Field Card
  • The mental model
  • Calibrate trust — don’t maximize it
  • Specify before you generate
  • Context hygiene
  • Workflow
  • Verify — where people get burned
  • When something feels off

§The AI Agent Operator’s Field Card

The difference between “shipped it” and “spent the afternoon cleaning up after it” with an AI coding agent is rarely the model. It’s the operator. Here is the whole discipline on one page. Free to read, copy, and share.

§The mental model

The agent is a fast, literal, tireless junior with no memory between sessions and zero stake in the outcome. It produces the work; you are accountable for it. That asymmetry never moves.

§Calibrate trust — don’t maximize it

Match scrutiny to stakes:

• Cheap if wrong + easy to verify → let it run, skim.
• Costly if wrong + hard to verify → read every line, or do it yourself.

Know which cell you’re in before you accept the diff.

§Specify before you generate

Name all four or the agent guesses — and it guesses generic:

1. Change — concretely (not “improve” / “handle”).
2. Scope — files in play, and what’s off-limits.
3. Constraints — conventions to match, what to preserve.
4. Done — a runnable check (a test, a command, an observable behavior).

And the highest-leverage trick: point at an existing example to follow.

§Context hygiene

• Starvation → feed it: name files, paste full errors.
• Pollution → reset: one session, one job.
• Long and muddy? Summarize → restart → reseed.
• Context is a workbench, not a filing cabinet.

§Workflow

• Plan first, read the plan, then let it write code.
• Decompose: one-sentence “done”, one-sitting diff.
• Commit each unit before the next.
• Two growing failed fixes in a row = thrash. Stop, revert, slow down.

§Verify — where people get burned

“The agent says it works” ≠ it works.

• Run it. Behavior outranks claims.
• Read the diff, not the summary.
• git diff --stat to catch silent scope creep.
• For agent-written tests, ask: would these fail if the code were broken?

§When something feels off

Run all three — one almost always finds it: specify earlier · verify harder · reset the context.

That’s the card. If it’s useful, the full Agent Operator’s Handbook expands every line above into a chapter — setup, specification, context, workflows, verification, the power tools (hooks, MCP, sub-agents, memory), a failure-mode catalogue, and a full recipes chapter. No fluff. Crypto checkout, instant delivery:

👉 https://are-history-targets-disposition.trycloudflare.com

#AI #coding #ClaudeCode #devtools