A recon-first workflow with Claude & Cursor

I treat AI agents the way I treat a junior pair: brilliant at execution, dangerous without a plan. So my workflow puts reconnaissance and review first, and keeps execution on a short leash.

The three phases

1. Recon — the agent reads the codebase and writes back what it found. No edits.
2. Plan — we agree on an approach, file by file, before a single line changes.
3. Execute — small, reviewable steps. I commit; the agent never pushes.

The agent proposes. I dispose. Git history stays mine.

The guardrail is one line in the agent’s instructions:

# Never push; I commit and push manually after reviewing.
git config --local receive.denyCurrentBranch true

Why it holds up

• Every change is a diff I actually read.
• Rollbacks are trivial because steps are small.
• The plan is the contract — drift gets caught early.

It’s slower per step and far faster overall, because I almost never throw work away.