AI Writing Workflow: Outline → Draft → Edit → Publish SOP

Updated: 2026-08-12

Why this guide

Stable AI writing is not “one prompt → 3,000 words.” It is outline → draft → edit → publish, each stage with inputs, prompts, and acceptance criteria. Treat ChatGPT (or your primary assistant) as a fast drafter; you own clarity, truth, and privacy. Official: ChatGPT, Claude, Gemini. Non-sensitive practice: domestic entry. Foundations: prompt tips.

Outcomes

• Fewer full rewrites
• Copy-ready prompts per stage
• Publish gates against hallucinations and leaks
• Consistent voice across sections

Four gates

Stage	You provide	Model delivers	Pass when
1 Outline	Topic, audience, goal, sources	Outline + risk spots	You approve structure; no invented chapters
2 Draft	Locked outline + material	Section prose	Covers outline; marks [VERIFY]
3 Edit	Draft + issues	Compress / tone / examples	Sounds like you
4 Publish	Final	Titles, blurb, checklist	Fact + sensitivity pass

Skipping the outline is the #1 failure mode.

Stage 1 — Outline (15–25 min)

Write five sentences: audience, outcome, bans, length, tone. Paste only sources you opened. Demand outline—not prose. Lock outline-v1.

Role: senior editor
Goal: outline for “{topic}”
Audience: { }; after reading they can: { }
Tone: { }; length target: {N}
Bans: {e.g. absolute medical claims}
Source bullets:
- {point + source}
Output: 3 title options; H2/H3 outline; per-section fact types to verify (no invented numbers); risk spots

Pass: every section has one job; no surprise “exclusive stats”; intro/outro connect.

Stage 2 — Draft (section by section)

One section at a time. Ban invented citations and stats. Stitch with transition-only passes.

Write only the section “{H2}” from the locked outline.
Use my material only; missing info → [NEED: …]
Numbers/dates/laws must be [VERIFY] or blank
No fake links/papers
~{N} words; same tone
Material: { } Prior section ending: { }
Output: section + verify list

Here are section drafts. Add transitions and minor heading tweaks only—do not rewrite arguments. Flag two redundancies.
Drafts: { }

Stage 3 — Edit (fixed order)

Order: correct → clear → voice. Do not chase “more literary” first.

Act as fact/structure reviewer. Check outline drift; list checkable claims; mark empty filler.
Annotate—do not rewrite into a new article.
Body: { }

Without changing claims: shorten long sentences; add one concrete example per section marked [EXAMPLE: replaceable]; cut repeats.
Output revised body.

Unify voice as “{e.g. calm, short sentences, few adjectives}”.
Keep terms; no slang pile-ups.
Show 5 before→after notes of what you changed.
Body: { }

Read aloud once. Fix awkward lines yourself before another “make it perfect” pass.

Stage 4 — Publish gates

• All [VERIFY] resolved with sources you opened
• Every outbound link loads and matches the claim
• No fake papers or expert quotes
• No client names/secrets (privacy)
• Title not stronger than evidence
• Meta description matches promises
• Regulated advice carries appropriate disclaimers

From the locked final, produce: 5 non-clickbait titles; ≤140-char summary; 3 FAQs answerable from the body (else say “not covered”).
Final: { }

Deep dive: AI fact-checking.

Material pack (teams)

# Article pack
Topic / Audience / Action goal / Bans /
Voice sample (1 paragraph of your past writing) /
Fact list with URLs you opened /
Redacted notes /
Reference links (structure only—do not copy) /
Deadline /

Genre tweaks

Type	Emphasize	Reduce
Tutorials	Copy-ready steps, screenshot slots	Vague motivation
Email/weekly	Conclusion first, clear owners	Metaphors
Proposals	Risks & assumptions listed	Absolute promises
Social	One hook + one action	Hashtag sludge

Also see office workflow.

Common mistakes

Mistake	Fix
“Write 3k words” in one shot	Lock outline; draft by section
Trust model citations	Only your opened sources
Edit out of order	Correctness before style
Multi-model paste wars	One primary; others annotate only
Skip sensitivity scan	Use privacy review prompt

Self-score (1–5)

Structure (one job/section) · Evidence (clickable checks) · Voice (one author) · Usefulness (executable) · Risk (no leaks/fake cites). Redo the weak stage—don’t reroll forever.

FAQ

Must I use ChatGPT? Is Claude better for longform?

Choose by task. Many prefer Claude for careful long edits; ChatGPT shines at iterative office drafts. Blind-test via selection guide.

Can I generate the full draft then edit?

In a rush, yes—higher error rate. Compromise: confirm intro+outline, then two sections per batch; still ban invented numbers.

How do I keep my voice?

Put a sample paragraph in the pack; specify observable rules in the voice pass.

Multi-author column?

Shared pack + outline template + publish checklist; authors may differ, gates should not.

Domestic entry for real posts?

Fine for non-sensitive drafting; account value stays on chatgpt.com. Access: China guide.

Related reading

• Prompt tips
• AI fact-checking
• AI privacy
• Getting started

Next reading

• AI study workflow
• Choose an AI assistant
• FAQ

Action path

Today: run Stage 1 and lock outline-v1. Tomorrow: sectional draft + three edit passes. Ship day: tick the publish checklist—facts and sensitivity before go-live. Help: OpenAI Help.