PowerUp Coaching — Living Knowledge Base

Insight: The Universal Audit Phrase — Errors, Omissions, Oversights, Duplications, Contradictions

6 min read

“I’ll say to it: audit for errors, omissions, oversights, duplications, contradictions, and — if applicable — areas of improvement that meet our criteria. So I’m telling it to tear down what’s not good, not only to add more stuff that is good.” — Lou, 2026-05-21

Session context: 2026-05-21_Mastermind — Dirk described his frustration with AI iteration loops that keep generating “improvements” that drag the work in circles. Lou’s answer was a specific phrase he keeps loaded as a programmed-in audit command. The phrase has six named failure modes and one explicit gate (“if applicable, improvements that meet our criteria”). It is one of the most copy-and-pasteable artifacts of the session.

Core Idea

When you ask an AI to “improve” something, you have given it permission to keep adding. Additive prompts produce additive outputs — more bullets, more clauses, more “considerations.” The work gets longer; it does not get better. The Universal Audit Phrase replaces improve with a list of named failure modes that the AI must actively search for and either fix or confirm absent:

  • Errors — claims that are factually wrong.
  • Omissions — things that should be there but are missing.
  • Oversights — implications of what was written that the writer did not notice.
  • Duplications — points made more than once, or near-restatements of earlier points.
  • Contradictions — places where the work disagrees with itself.

Then a sixth pass — explicitly gated:

  • Improvements — but only those that meet our criteria.

The gating clause is the lever. It forces the AI to check against an existing standard before adding anything, which prevents the “let me suggest 10 more things” failure mode. If the criteria are clear, most of the model’s suggestion impulse gets filtered out at this gate. If the criteria are unclear, the model surfaces that fact — which is itself useful, because it tells you the standard is not actually defined.

This is conceptually adjacent to Insight - Always Audit Your Plan Before You Build — The 18-Problem Discovery (audit-before-execute as a standing discipline) and Insight - Prevent AI Drift by Treating System Prompts as Living Constraints (constraints as the active guard against drift). What is new is that this is a single phrase you can paste anywhere — into a CLAUDE.md, into the middle of a session, into a slash command — and it consistently produces work that is tightened rather than extended.

Why This Matters for Knowledge Entrepreneurs

The default AI failure mode in creative and strategic work is sycophantic accretion. The model wants to be helpful. “Improve this” is interpreted as “add value,” and added value almost always means more words, more options, more frameworks. After three rounds of “improvements,” the document is twice as long and half as sharp.

The audit phrase inverts the default. By naming what could be wrong instead of asking what could be better, you redirect the model from generation back to critique. Critique is where AI judgment is most useful and least dangerous — the model is good at spotting contradictions, duplications, and missing pieces in text it has just been shown. It is much worse at deciding what would constitute a better version, because “better” is undefined unless you defined it.

For knowledge entrepreneurs producing content, frameworks, and client deliverables under time pressure, this phrase is the single most leverage-positive copy-paste in this vault. It costs nothing to use. It produces work that is shorter, denser, and more internally consistent. And because the failure modes are named, the model’s output is auditable in turn — you can see whether it actually found contradictions or just made up new ones to demonstrate effort.

Practical Application

Copy this phrase into your CLAUDE.md, your .claude/rules/, or whichever system prompt the AI sees every session:

Audit this for errors, omissions, oversights, duplications, contradictions, and — if applicable — areas of improvement that meet our criteria.

Use it at three specific moments:

  1. After any AI-generated draft. Before you read the draft, run the audit on it. The AI will surface its own weak points more honestly when prompted critically than when prompted for praise.
  2. Mid-conversation, before a “let’s keep going” instinct. If a session has produced 5+ rounds of output, run the audit before continuing. Half the time the audit reveals the work is already converged and the additional rounds would be sycophantic accretion.
  3. At the end of any document you are about to ship. Final pre-publication critique. The audit catches the late-stage damage that revisions tend to introduce.

The “meet our criteria” clause is load-bearing. If your criteria are not written down, write them down first — even a one-line “the audience for this is a solo knowledge entrepreneur who needs to implement it this week” gates 80% of the noise.

Coaching question: “If I had to choose between asking the AI to ‘improve’ this or to ‘audit it for what’s wrong with it,’ which one would actually move the work forward — and which one am I defaulting to?”

Evolution Across Sessions

Builds on Insight - Always Audit Your Plan Before You Build — The 18-Problem Discovery (2026-04-30), which established that audit-before-execute consistently surfaces an outsized number of real problems. The new development is the exact lexical phrase Lou has reduced this discipline to, and the placement of the explicit meet-our-criteria gate on the “improvements” clause. Prior insights talked about auditing as a practice; this one names the words that make the practice repeatable.

Source

  • 2026-05-21_Mastermind (Lou — in response to Dirk’s frustration with AI iteration loops; quoting the exact phrase he has programmed into his CLAUDE.md)

Graph View

Backlinks