Andy University coding agents for real beginners

Guide Note 5

Step Back And Review Your Setup

Step outside your working setup and find the blind spots it cannot see while you are using it.

Open the short practice promptBack To Start Here

Short Version

What this means if you are still new.

Beginners do not need to obsess over formal reviews on day one, but they do need the idea early: comfortable setups can still be wrong.

Example

Sales example: if your agent always confirms your preferred narrative about a prospect or strategy, you need a review session that tries to break your frame.

Open a short practice prompt Use this to feel the lesson without reading the full notes first.

Pasteable Prompt

Try this today

I want you to review my current way of using an agent.

Your job is not to flatter the setup.
Your job is to find where it is actually wrong, stale, or too comfortable.

Give me three concrete corrections.

This is the shortest possible action that lets you feel the lesson in practice.

Longer Notes

Open the full guide notes only when you want more depth.

This keeps Andy's original longer write-up available without forcing all of it onto a first-time visitor.

Open the full notes for guide note 5 Andy calls this a meta-layer audit. This is the longer original version.

You've built a constitution. You run synthesis sessions. You close your sessions. You run quality loops. And some of this might be subtly wrong — but you can't see it, because you're running the system from inside it.

The same session that operates your constitution is running on your constitution's priors. It's optimizing for what your constitution says is good. If your constitution has a blind spot, the session operating from it will confirm the blind spot, not challenge it.

This is not a failure of AI. It's a structural property of any system that runs on its own rules.

The Technique

Quarterly, or after major milestones: run a session that is architecturally separate from your normal operating mode.

The audit session has one explicit mandate: find where the primary system is wrong.

What makes it architecturally separate:

  • Different context loaded at the start. Instead of your constitution, load the audit template below.
  • The explicit instruction to the AI: "Your job is to find where my working methodology is wrong. Not where it could be improved — where it is actually wrong."
  • You cannot be in working mode and audit mode simultaneously. Declare the audit explicitly.

The Audit Template

Session mandate: find where the primary system is wrong.

Check 1 — Confirmation patterns:
  Review the last 10 synthesis sessions or working sessions.
  Did they produce outputs consistent with what the constitution predicts I value?
  If yes: is that because the constitution is right, or because sessions confirm it?

Check 2 — Stale assumptions:
  Look at the decision logic section of the constitution.
  What premises is it built on? Are those still true?
  [List assumptions. Mark stale/current.]

Check 3 — Untested constraints:
  What has been treated as non-negotiable without testing?
  [List. Note which were tested and which were assumed.]

Check 4 — Avoidance patterns:
  Where have I consistently found reasons not to proceed?
  What gets deferred repeatedly without resolution?

Output: 3-5 specific updates to the constitution.
  Not aspirational improvements — actual corrections. Things that were wrong.

The Quality Test

A useful audit produces at least one finding that makes you slightly uncomfortable. If the audit was entirely comfortable — "everything is basically working, maybe I should add more detail here" — it was probably confirming rather than auditing.

The test: did the audit change your constitution in a way you wouldn't have predicted before running it?

When to Run It

Minimum: quarterly. Also after: a major context change (new job, new project, significant shift in what you're working on), after running the synthesis engine on a question where your constitution seemed to produce the wrong framing.

Do not run it too frequently. A system needs time to prove itself before you destabilize it. Monthly audits of a two-week-old constitution produce noise, not signal.

Practice

Run your first meta-layer audit using the template above.

Completion criteria: An audit document exists with at least 3 findings. At least one finding is something that was subtly wrong — not "this could be improved" but "this was actually not right."

What You've Built

After completing all five lessons, you have:

  1. A working constitution (in active use, being updated)
  2. A session close habit (at least 5 sessions closed)
  3. A synthesis sequence (at least one complete 5+ pass telescope)
  4. An active quality loop (at least one cycle complete)
  5. One meta-layer audit (constitution updated with findings)

That's an AI operating model. Not theoretical — working, with artifacts to prove it. The methodology compounds from here.