2026-05-01

When a one-pass review beats a three-pass review

I built three perfection-pass commands. The expensive one isn't always the right call. Here's the budget I use.

#code-review#agents#perfection-passes#claude-code

The lineup

Three commands in my Claude Code setup, all do code refinement, in increasing cost:

  • /critique — single-pass 3-agent critique. ~$0.10. ~3 minutes.
  • /pp2p — two perfection passes plus polish. ~$0.40. ~10 minutes.
  • /pp3p — three perfection passes plus polish, includes Codex adversarial review. ~$1.20. ~25 minutes.

The temptation, every time, is to reach for /pp3p on everything. Don't.

When /critique is enough

  • Drafts I'm going to revise myself anyway
  • Internal-only scripts that don't ship to production
  • Documentation
  • Anything where the cost of a missed bug is "I fix it later"

/critique runs three agents in parallel — different angles, no cross-pollination. You get three perspectives in one read. Most code I write needs exactly that.

When /pp2p is the right call

  • Features that touch the database
  • Anything that crosses a trust boundary (API auth, file uploads, user input)
  • Public-facing UI changes
  • Anything I'd be embarrassed if a customer saw broken

The two-pass cycle catches what one pass misses. The polish pass catches what perfection misses. About 90% of "real" code I ship gets /pp2p.

When /pp3p earns its slot

  • Auth code
  • Payment code
  • Database migrations on tables with real data
  • Anything where the cost of a bug is "I can't undo this"

The third pass + Codex adversarial review catches the bugs the other passes missed because they were thinking like me. Codex thinks differently. That's the unlock.

The budget

Rough monthly spend on these passes: about $40. Bugs caught: probably $4,000 worth, if I count "embarrassment cost" plus "incident response time."

The trap is using /pp3p on something that should have been /critique. That's $1.10 wasted per use, plus 22 minutes of wall-clock time. Do it 30 times and you've burned a workday.

What you can steal

If you're using AI code review at all, build the cost ladder explicitly:

  1. Cheap pass for drafts and internal scripts
  2. Medium pass for production code
  3. Expensive adversarial pass for irreversible changes

Then DECIDE which tier each commit belongs to BEFORE you run the review. The discipline is in the assignment, not in the tool.

I get this wrong about once a week — reach for /pp3p on something that didn't need it. The fix is to slow down by 10 seconds and ask: "What's the cost of a bug here?" Then pick the tier that matches.