Writing

Self-Critique Line Edit Pass

A two-phase editing prompt: the model first diagnoses concrete problems in your draft against named criteria, then rewrites only the passages it flagged.

Prompt template

You are a line editor. You will work in two phases and must not begin
phase two until phase one is complete.

## Draft
{{DRAFT}}

## Voice to preserve
{{VOICE_NOTES}}

## Phase 1 — Diagnose
Read the draft and produce a numbered list of specific problems. For each
problem give: the exact quoted text, the category, and one sentence on why
it fails. Use only these categories:
- BURIED: the point arrives after throat-clearing
- VAGUE: an abstract claim where a concrete noun or number belongs
- PADDING: words that carry no information
- UNSUPPORTED: an assertion the draft never backs up
- RHYTHM: consecutive sentences of near-identical length or structure
- HEDGE: qualifiers that drain a claim without adding accuracy

Find between 5 and 12 problems. If the draft genuinely has fewer, say so
rather than inventing issues.

## Phase 2 — Repair
Rewrite ONLY the passages you quoted in phase 1. Present each as:
Before: (original)
After: (revision)
Why: (one line)

## Constraints
- Do not rewrite passages you did not flag.
- Do not change the argument, add claims, or introduce facts not in the draft.
- Preserve the voice described in {{VOICE_NOTES}}; if a fix would flatten that
  voice, flag it and leave the text alone.
- Never replace a specific detail with a general one.
- Do not output a full clean version of the draft.

Why this works

This is a self-critique loop with a fixed diagnostic taxonomy. Asking a model to 'improve' a draft produces homogenized prose because it optimizes toward its own default register. Forcing an explicit critique pass first, with a closed category list, converts a vague aesthetic task into a classification task the model does well — and restricting phase two to the flagged spans is what preserves your voice everywhere else.

Example output

Example

## Phase 1 — Diagnose
1. "In the modern data landscape, organizations are increasingly finding
   that..." — PADDING. Eleven words before the sentence begins.
2. "significantly improved performance" — VAGUE. No baseline, no metric,
   no magnitude.
3. "It could be argued that this may sometimes be a factor" — HEDGE. Three
   qualifiers stacked on one weak claim.

## Phase 2 — Repair
Before: In the modern data landscape, organizations are increasingly
finding that query latency matters.
After: Query latency matters.
Why: The preamble carried no information the next sentence did not.

Before: significantly improved performance
After: cut p99 latency roughly in half
Why: Replaced an abstraction with the measured effect.

Usage tips

Paste a draft of under about 1,200 words. Beyond that, models start sampling rather than reading closely, and the diagnosis gets thin toward the end. Split long pieces by section.

Fill VOICE_NOTES with concrete rules, not adjectives. “Short sentences, no semicolons, first person plural, occasional sentence fragments for emphasis” works. “Punchy and conversational” does not.

Why two phases beat one

A single-pass edit request gives the model no reason to justify its changes, so it rewrites everything and you lose the ability to audit what happened. Diagnosis-then-repair produces a diff you can accept or reject item by item — and rejecting item three does not disturb items one and two.

Browse library

Technical Blog Post Outline

Turns a topic and audience into a structured outline with per-section word budgets, a stated thesis, and an explicit list of what the post will not cover.

Bullet Notes to Voice-Matched Draft

Converts rough notes into a first draft that imitates a supplied writing sample, with strict rules against inventing facts the notes do not contain.

The AI signal without the hype — new models, tools worth your time and what actually shipped. One email, no vendor pitches.

Choosing an AI stack for your team?

We help companies pick the right AI tools, wire them into existing systems and avoid the ones that quietly do not scale. Tell us what you are trying to build and we will tell you what we would use — no charge for the conversation.