PowerUp Coaching — Living Knowledge Base

The Six Words That Replace 'Improve This' — A Universal Audit Phrase for AI Work

4 min read

Topic

Why “improve this” produces longer, weaker AI output — and the exact six-word audit phrase that produces shorter, tighter, more honest output instead.

Target Reader

Knowledge entrepreneurs who use Claude/ChatGPT daily for content, frameworks, client deliverables, and decision support — and who have noticed that iterative “improvement” loops tend to produce more output, rarely better output.

The Fear / Frustration / Want / Aspiration

The frustration: AI iteration loops that keep adding bullets, considerations, and clauses until the document is twice as long and half as sharp. The fear that you are publishing sycophantic accretion. The want: a single phrase that tightens AI output instead of bloating it. The aspiration: AI as a critical reviewer, not a sycophantic helper.

Before State

The reader uses “improve this,” “make it better,” “polish this” — and watches the work get longer with each iteration. They suspect the AI is adding noise but cannot consistently catch it. They lack a vocabulary for getting the AI into critique mode.

After State

The reader has the audit phrase pasted into their CLAUDE.md and uses it reflexively after every draft. AI output becomes auditable on six named axes. Iteration produces tighter work, not longer work. They stop using “improve this” entirely.

Narrative Arc

Open with Dirk’s frustration on the AIM call — the spiral of Claude Chat → Cowork → Code where every pass adds new “improvements” and nothing ever finishes. Name the failure mode: additive prompts produce additive output. Reveal the phrase. Walk through each of the six failure modes (errors, omissions, oversights, duplications, contradictions, gated improvements) and show what each catches that “improve this” misses. Close with the practical implementation — paste it into CLAUDE.md, use at three specific moments, codify as a slash command.

Core Argument

“Improve this” is a permission slip for the AI to keep adding. The Universal Audit Phrase — “audit this for errors, omissions, oversights, duplications, contradictions, and if applicable, areas of improvement that meet our criteria” — replaces an open-ended additive prompt with a closed list of named failure modes. The “meet our criteria” gate on the improvements pass is the load-bearing word that prevents the model from inventing new standards mid-audit.

Key Evidence / Examples

  • Dirk’s looping-iteration frustration on the 2026-05-21 AIM call — the exact use case the phrase was designed to solve.
  • The six failure modes broken out concretely: an error is wrong; an omission is missing; an oversight is unnoticed implication; a duplication is restated content; a contradiction is internal disagreement; an improvement is additive but gated.
  • The “improve this” pathology: most AI editors are trained to be helpful, which means adding value, which means adding words.
  • The CLAUDE.md placement pattern from Lou’s setup — the phrase as a standing constraint, not a one-shot prompt.
  • Cross-reference to Insight - Always Audit Your Plan Before You Build — The 18-Problem Discovery: the broader audit-before-execute discipline this phrase operationalises.

Proposed Structure (5–7 beats)

  1. The improvement spiral. Open with Dirk’s loop. Set up the failure mode of additive prompting.
  2. What “improve” actually means to an LLM. “Add value” = “add words.” Why this is structural, not a model defect.
  3. The phrase. Reveal it in full. The six failure modes plus the gated improvement clause.
  4. Why each word does work. Walk the six failure modes one by one. Show what each catches that “improve this” misses.
  5. The load-bearing gate. Why “meet our criteria” is the most important phrase in the audit. The gate that prevents the model from inventing new standards.
  6. Three places to use it. After any draft. Mid-conversation before “let’s keep going.” Before shipping anything.
  7. How to codify. CLAUDE.md placement. Slash command wrapper. Re-running the audit until output stabilises.

Editorial Notes

Tactical-craft register. The reader should leave with the exact phrase in their clipboard and a clear instruction to paste it into their CLAUDE.md. Resist the temptation to elaborate the phrase — its power is in being short enough to memorise. The article should be the shortest of the briefs in this batch; the artifact carries the weight.

The opening must not be theory-first. Open with a concrete failure mode in the wild — Dirk’s loop or a reader-relatable parallel.

Next Step

  • Approved for drafting
  • Needs revision
  • Deprioritised

Graph View