Draft Review Revise Loop
Parent: Writing and Documentation · researched 2026-05-29T19:26:36.137Z· 6 sources · 10 concepts · skill draft-review-revise-loop
Purpose: a meta-skill that coordinates the three modes of writing work — generating,
Draft → Review → Revise Loop
- Purpose: a meta-skill that coordinates the three modes of writing work - generating, [source]
- evaluating, and improving - in discrete passes with explicit handoff points. The loop [source]
- is not infinite. It has stop conditions. [source]
- 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. [source]
- Output shape per iteration: [source]
- Review findings list - each finding on its own line with severity label: [Critical], [High], [Medium], or [Low], plus a one-line description [source]
- Revision summary (what changed, why, keyed to finding IDs) [source]
- Stop-condition check (which condition applies, or "continuing to iteration N+1") [source]
- After completing each loop, re-read the output and confirm it addresses the user's stated goal before delivering it. [source]
The Three-Pass Discipline
Pass 1 — Draft
- Mode: generative. Your job is to get ideas out, not to write well. [source]
- No editing while drafting. Resist fixing sentences mid-flow. Let ugly prose stand. [source]
- Anne Lamott calls this the "shitty first draft" - a permission slip to write badly so you [source]
- can write at all. The draft's only job is to exist. [source]
- Mark gaps, don't plug them mid-draft. Use [TK] (to come) or `[GAP: describe what's [source]
- missing]` for unknowns. Stopping to research kills momentum. Fill gaps in Pass 3. [source]
- **Any [TK] or [GAP:] still present at the end of Pass 3 is a Critical finding - the [source]
- document is not shippable until all markers are resolved or explicitly scoped out.** [source]
- Time-box. One Pomodoro - 25 minutes - per major section. If the clock runs out, stop [source]
- mid-sentence and move on. Unfinished prose is not a crisis; it is raw material. [source]
- Work from an outline, not from instinct. Even a three-bullet skeleton prevents the [source]
- most common draft failure: a conclusion that argues against the introduction. [source]
- See the Skill Handoff Map at the end of this document for the full per-phase routing table. [source]
Pass 2 — Review
- Mode: evaluative. You are now a reader, not the author. [source]
- Read the whole thing once before commenting. No inline notes on the first pass. [source]
- You need to know where the document goes before you judge any sentence in it. [source]
- Separate findings by type: [source]
- Missing: a claim is made but the evidence is absent; a step is assumed but not [source]
- stated; a section promised in the introduction does not appear. [source]
- Wrong: a fact is incorrect; an argument contradicts itself; a structural pattern [source]
- Could be better: prose is clear enough but could be sharper, shorter, or better [source]
- Name the framework you are reviewing against. Do not review in a vacuum. Pick one [source]
- BLUF - does the bottom line appear in the first sentence? [source]
- MECE - are sections mutually exclusive and collectively exhaustive? [source]
- Minto Pyramid - does the argument flow deductively from a single governing idea? [source]
- PARA - is the document in the right context (project, area, resource, or archive)? [source]
- SCQA - is the Situation / Complication / Question / Answer arc intact? [source]
- Severity-label every finding: Critical (document is wrong or misleading), High [source]
- (missing material the audience needs), Medium (structure or argument weakness), Low [source]
- (style preference, nice-to-have). [source]
- Do not fix during review. Record findings. Mixing evaluation and correction corrupts [source]
- both modes - you stop seeing what is actually there. [source]
Pass 3 — Revise
- Mode: surgical. Apply the review findings against a priority threshold. [source]
- Address every Critical and High finding. These are not optional. If a Critical [source]
- finding cannot be resolved, the document is not ready to ship - restart or descope. [source]
- Batch Medium findings. Group related Medium findings and address them in one pass [source]
- through the affected section. Avoid thrashing individual sentences. [source]
- Skip Low findings unless trivial. A Low finding that takes thirty seconds to fix [source]
- is worth taking. A Low finding that requires restructuring a paragraph is not worth [source]
- taking on the first revision cycle - bank it for the next iteration if it recurs. [source]
- Do not introduce new content during revision. If revision uncovers a missing [source]
- section, that section goes back to draft mode (Pass 1), not inline into the revision [source]
- pass. Keep the modes clean. [source]
- Log what changed and why. One-line entry per finding resolved: [source]
- [H] §3 - added missing latency data; source: internal benchmark doc. [source]
- See the Skill Handoff Map at the end of this document for routing after revision. [source]
Stop Conditions
The 24-Hour Cooling-Off Rule
- Stephen King recommends letting a draft rest before revising. The principle: you cannot [source]
- read what you wrote - you read what you meant to write. Distance corrects that. [source]
- For documents under 500 words: 1–2 hours is sufficient. [source]
- For documents 500–2,000 words: overnight. [source]
- For documents over 2,000 words: 24 hours minimum. [source]
- If the deadline prevents cooling off, use the fresh eyes pattern instead. [source]
The Fresh Eyes Pattern
- If you cannot wait, get a different reviewer for iteration N+1 than for iteration N. [source]
- The first reviewer has already accommodated your argument; they will find it harder to [source]
- see what is still missing. A second reviewer reads with no accommodation debt. [source]
- Applied to Claude: ask a second agent session (or invoke document-critique with a cold [source]
- context) rather than asking the same session that helped draft to also review. [source]
Per-Iteration Log
- Keep a running log for every document that goes through more than one cycle. [source]
Anti-Patterns
- Editing while drafting. Kills generative momentum. You end up with the first three [source]
- paragraphs polished and the rest unwritten. Lamott diagnosed this decades ago; the fix [source]
- is the same: write badly first. [source]
- Infinite polishing. Zinsser: "Rewriting is the essence of writing well - but there [source]
- is a difference between rewriting that improves and rewriting that avoids." Apply the [source]
- convergence and time-box stop conditions ruthlessly. [source]
- Five-stakeholder review before any iteration. Gathering input from five people on a [source]
- first draft means gathering five reactions to a document that will change dramatically [source]
- anyway. Get to iteration 2 before widening the reviewer pool. [source]
- No time-box on review. A review pass without a time-box expands to fill all available [source]
- attention. Cap each review pass at half the draft time or 30 minutes, whichever is less. [source]
- Skipping the framework. Reviewing without a named framework produces impressionistic [source]
- feedback ("this feels off") that is hard to act on. Name the lens before you start. [source]
When to Bail Entirely
- Iteration is not always the answer. Stop the loop and start over when: [source]
- Scope changed. The document is now answering a different question than it started [source]
- with. Revising toward a moving target produces incoherent documents. [source]
- Audience changed. A document written for engineers and redirected to executives [source]
- needs a new draft, not another revision pass. [source]
- Premise was wrong. The core argument has been disproven or superseded. No amount [source]
- of revision fixes a document whose central claim is no longer true. [source]
- In all three cases: close the log, open a new one, and return to Pass 1. The old draft [source]
- is not wasted - it is a source of raw material and a record of what the argument was. [source]
Children
- Shitty first draft (Lamott) (frontier)
- Named-framework review (BLUF, MECE, Minto, SCQA) (frontier)
- Severity-ranked findings (Critical / High / Medium / Low) (frontier)
- Hard stop (3 iterations) (frontier)
- Soft stop (no medium-or-higher findings) (frontier)
- Convergence stop (5% word-count delta) (frontier)
- Time-box stop (50% draft effort) (frontier)
- 24-hour cooling-off rule (frontier)
- Fresh eyes pattern (frontier)
- Bail conditions (scope / audience / premise changed) (frontier)
Frontier under this node: 24-hour cooling-off rule, Bail conditions (scope / audience / premise changed), Convergence stop (5% word-count delta), Fresh eyes pattern, Hard stop (3 iterations), Named-framework review (BLUF, MECE, Minto, SCQA), Severity-ranked findings (Critical / High / Medium / Low), Shitty first draft (Lamott), Soft stop (no medium-or-higher findings), Time-box stop (50% draft effort)