Draft Review Revise Loop

Draft → Review → Revise Loop

Purpose: a meta-skill that coordinates the three modes of writing work — generating, evaluating, and improving — in discrete passes with explicit handoff points. The loop is not infinite. It has stop conditions.

When invoked by Claude: produce a structured pass log for each iteration — findings by severity, actions taken, and which stop condition ended the loop. Do not silently revise without showing the iteration record. If the user provides a document, run the loop against it; if they describe a document, ask for the text before proceeding.

Output shape per iteration:

1. Review findings list — each finding on its own line with severity label: [Critical], [High], [Medium], or [Low], plus a one-line description
2. Revision summary (what changed, why, keyed to finding IDs)
3. Stop-condition check (which condition applies, or “continuing to iteration N+1”)

After completing each loop, re-read the output and confirm it addresses the user’s stated goal before delivering it.

The Three-Pass Discipline

Writing fails in two ways: shipping too early (underdeveloped argument, missing structure) or never shipping (infinite-polish). The three-pass loop prevents both by separating incompatible cognitive modes and capping iterations before they become procrastination.

Pass 1 — Draft

Mode: generative. Your job is to get ideas out, not to write well.

Rules:

• No editing while drafting. Resist fixing sentences mid-flow. Let ugly prose stand. Anne Lamott calls this the “shitty first draft” — a permission slip to write badly so you can write at all. The draft’s only job is to exist.
• Mark gaps, don’t plug them mid-draft. Use [TK] (to come) or [GAP: describe what's missing] for unknowns. Stopping to research kills momentum. Fill gaps in Pass 3. Any [TK] or [GAP:] still present at the end of Pass 3 is a Critical finding — the document is not shippable until all markers are resolved or explicitly scoped out.
• Time-box. One Pomodoro — 25 minutes — per major section. If the clock runs out, stop mid-sentence and move on. Unfinished prose is not a crisis; it is raw material.
• Work from an outline, not from instinct. Even a three-bullet skeleton prevents the most common draft failure: a conclusion that argues against the introduction.

See the Skill Handoff Map at the end of this document for the full per-phase routing table.

Pass 2 — Review

Mode: evaluative. You are now a reader, not the author.

Rules:

• Read the whole thing once before commenting. No inline notes on the first pass. You need to know where the document goes before you judge any sentence in it.
• Separate findings by type:
  • Missing: a claim is made but the evidence is absent; a step is assumed but not stated; a section promised in the introduction does not appear.
  • Wrong: a fact is incorrect; an argument contradicts itself; a structural pattern is violated.
  • Could be better: prose is clear enough but could be sharper, shorter, or better ordered.
• Name the framework you are reviewing against. Do not review in a vacuum. Pick one and be explicit:
  • BLUF — does the bottom line appear in the first sentence?
  • MECE — are sections mutually exclusive and collectively exhaustive?
  • Minto Pyramid — does the argument flow deductively from a single governing idea?
  • PARA — is the document in the right context (project, area, resource, or archive)?
  • SCQA — is the Situation / Complication / Question / Answer arc intact?
• Severity-label every finding: Critical (document is wrong or misleading), High (missing material the audience needs), Medium (structure or argument weakness), Low (style preference, nice-to-have).
• Do not fix during review. Record findings. Mixing evaluation and correction corrupts both modes — you stop seeing what is actually there.

Pass 3 — Revise

Mode: surgical. Apply the review findings against a priority threshold.

Rules:

• Address every Critical and High finding. These are not optional. If a Critical finding cannot be resolved, the document is not ready to ship — restart or descope.
• Batch Medium findings. Group related Medium findings and address them in one pass through the affected section. Avoid thrashing individual sentences.
• Skip Low findings unless trivial. A Low finding that takes thirty seconds to fix is worth taking. A Low finding that requires restructuring a paragraph is not worth taking on the first revision cycle — bank it for the next iteration if it recurs.
• Do not introduce new content during revision. If revision uncovers a missing section, that section goes back to draft mode (Pass 1), not inline into the revision pass. Keep the modes clean.
• Log what changed and why. One-line entry per finding resolved: [H] §3 — added missing latency data; source: internal benchmark doc.

See the Skill Handoff Map at the end of this document for routing after revision.

Stop Conditions

The loop ends when the first applicable condition is met — earliest wins.

If you reach the hard stop without reaching soft stop, ship the best available version with a known-issues note, or escalate the scope problem. Infinite iteration is not a quality strategy.

The 24-Hour Cooling-Off Rule

Stephen King recommends letting a draft rest before revising. The principle: you cannot read what you wrote — you read what you meant to write. Distance corrects that.

For documents under 500 words: 1–2 hours is sufficient. For documents 500–2,000 words: overnight. For documents over 2,000 words: 24 hours minimum.

If the deadline prevents cooling off, use the fresh eyes pattern instead.

The Fresh Eyes Pattern

If you cannot wait, get a different reviewer for iteration N+1 than for iteration N. The first reviewer has already accommodated your argument; they will find it harder to see what is still missing. A second reviewer reads with no accommodation debt.

Applied to Claude: ask a second agent session (or invoke document-critique with a cold context) rather than asking the same session that helped draft to also review.

Per-Iteration Log

Keep a running log for every document that goes through more than one cycle.

## Revision log — [Document title]

### Iteration 1
- Draft completed: [date/time]
- Reviewed by: [agent or name], framework: [BLUF / MECE / etc.]
- Findings: 2 Critical, 3 High, 4 Medium, 1 Low
- Revisions applied: [brief list]
- Stop check: High findings remain — continue

### Iteration 2
- Draft completed: [date/time]
- Reviewed by: [different reviewer], framework: [BLUF / MECE / etc.]
- Findings: 0 Critical, 0 High, 2 Medium, 3 Low
- Revisions applied: [brief list]
- Stop check: No Medium-or-higher findings — SOFT STOP, ship

Anti-Patterns

Editing while drafting. Kills generative momentum. You end up with the first three paragraphs polished and the rest unwritten. Lamott diagnosed this decades ago; the fix is the same: write badly first.

Infinite polishing. Zinsser: “Rewriting is the essence of writing well — but there is a difference between rewriting that improves and rewriting that avoids.” Apply the convergence and time-box stop conditions ruthlessly.

Five-stakeholder review before any iteration. Gathering input from five people on a first draft means gathering five reactions to a document that will change dramatically anyway. Get to iteration 2 before widening the reviewer pool.

No time-box on review. A review pass without a time-box expands to fill all available attention. Cap each review pass at half the draft time or 30 minutes, whichever is less.

Skipping the framework. Reviewing without a named framework produces impressionistic feedback (“this feels off”) that is hard to act on. Name the lens before you start.

When to Bail Entirely

Iteration is not always the answer. Stop the loop and start over when:

• Scope changed. The document is now answering a different question than it started with. Revising toward a moving target produces incoherent documents.
• Audience changed. A document written for engineers and redirected to executives needs a new draft, not another revision pass.
• Premise was wrong. The core argument has been disproven or superseded. No amount of revision fixes a document whose central claim is no longer true.

In all three cases: close the log, open a new one, and return to Pass 1. The old draft is not wasted — it is a source of raw material and a record of what the argument was.

Skill Handoff Map