Most “expand this article” prompts produce padding: more words, no more value. That is exactly what Google now penalizes. As of June 2026, the Helpful Content system is baked into the core ranking algorithm and rewards information density over length — a tight 700-word page routinely outranks a bloated 3,000-word guide that buries the answer. So the goal is not “make it longer.” The goal is to find where the draft is thin (a claim with no evidence, a paragraph with no example, a section that skips the mechanism) and fill only those gaps.
These 12 prompts do that. Each one targets a specific weakness instead of asking the model to “write more.”
TL;DR
- Word count is not a ranking factor. Expand only where the draft is genuinely thin; cut anything that does not carry information.
- Run expansion in order: find thin sections (#1) → add evidence (#2) and examples (#3) → add depth and counter-points (#4–#6) → bloat audit (#12).
- For long drafts, use a model with a real long context: Claude Opus 4.7, Sonnet 4.6, and Gemini 3.1 Pro all hold 1M tokens. ChatGPT Plus only exposes ~320 pages of in-app context (full 1M needs the $200 Pro tier).
- Never let the model invent statistics. In 2026, hallucinated citations still appear in over 30% of chatbot answers in research contexts — fact-check every number.
- AI drafts passes 1 and 2; a human ships pass 3.
Who this is for
Content marketers building pillar pages for SEO, bloggers turning a quick draft into a definitive guide, and technical writers fleshing out a skeleton outline. If you write 1,000-plus-word pieces and keep hearing “this feels thin,” these prompts are aimed at you.
When not to use these prompts
Skip them when the underlying thesis is weak. Expansion amplifies whatever is already there — including a bad argument. And do not expand to hit an arbitrary word target. If 900 words fully answers the question, ship 900 words. The fastest way to fail a Helpful Content review is to inflate a complete answer into a padded one.
Pick the right model for the job
Expansion needs the model to hold your entire draft plus any reference material in memory at once, then reason across it. That makes context window and instruction-following the two traits that matter most. Pricing and limits below are as of June 2026.
| Model | Context window | API cost ($/1M in→out) | Best for |
|---|---|---|---|
| Claude Opus 4.7 | 1M tokens | $5 → $25 | Synthesis, technical depth, staying consistent across a long draft |
| Claude Sonnet 4.6 | 1M tokens | $3 → $15 | The everyday workhorse — most expansion runs |
| Gemini 3.1 Pro | 1M tokens | $2 → $12 | Cheapest 1M-token option; strong creative variation |
| GPT-5.5 (ChatGPT Plus) | ~320 pages in-app | $5 → $30 | Fresh angles and varied prose; full 1M context only on $200 Pro |
Practical rule: for a draft over ~5,000 words plus sources, paste everything into Claude or Gemini rather than ChatGPT Plus, which will silently truncate older context. For voice and “make this sound less robotic” passes, GPT-5.5 tends to read most alive.
The expansion checklist
Work the prompts in this sequence rather than firing them at random:
- Find the thin sections (prompt #1) — get a ranked gap list first.
- Add evidence (#2) and examples (#3) to the top gaps.
- Add depth: mechanism (#5), counter-points (#4), objections (#6).
- Add structure where a claim needs proof: comparison table (#7), step list (#8).
- Add an original angle (#11) so the piece is not a restatement.
- Run the bloat audit (#12) last, after everything else has landed.
12 copy-ready prompt templates
Each prompt assumes you paste the draft where you see [draft]. Replace bracketed placeholders with your own text — do not leave the brackets in.
1. Find thin sections
Read this draft: [draft]. Identify thin sections: (a) claims without
evidence, (b) paragraphs without an example, (c) sections under 100 words
on important points, (d) bullets that need expansion. Output a list ordered
by reader impact, not by document order.Swap: [draft]. Start here every time — it tells you where to spend effort.
2. Claim → evidence
For each unsupported claim in this draft, add one sentence of evidence: a
statistic with a named source, a named example, or a concrete first-hand
observation. Skip vague "studies show" — name the study or drop the claim.
Flag any claim you cannot support so I can verify it myself.The “flag what you cannot support” line matters: it surfaces the spots where the model would otherwise fabricate.
3. Example layer
For each abstract concept in this draft, add one concrete example: a named
tool, a real number, a real timeline. Aim for one example per major
paragraph. Do not reuse the same example for two different points.4. Counter-point section
Add a "when this doesn't apply" section. Acknowledge 2-3 legitimate cases
where the article's thesis fails or has limits. Tone: confident, not
defensive. This builds trust, so do not hedge it into mush.5. Depth via “what’s actually happening”
Pick the 2 most important concepts in this article. For each, add a
150-200 word "under the hood" passage: the mechanism, why it works that
way, and the thing most readers get wrong about it. Skip concepts that are
already explained well.6. Reader objection responses
List 3 objections a skeptical reader would raise against this draft. Add a
one-paragraph response to each, woven into the section where the objection
arises — not collected at the end. Readers object in context, so answer in
context.7. Comparison expansion
This article compares [X] and [Y] at a high level. Add a comparison table
with 3-5 rows of concrete differences (cost, learning curve, ceiling,
time-to-result), plus a one-line "pick X when… / pick Y when…" under it.8. Step-by-step from a high-level claim
The article says "you can do [X]" but never shows how. Add a 5-step
"how to actually do this" sidebar. Each step = action verb + the outcome it
produces. No generic advice like "plan carefully" or "stay consistent."9. Glossary inline expansion
Identify jargon used without definition. For each term, add an inline
parenthetical definition (15 words or fewer) the first time it appears.
Do not build a glossary at the end — readers don't cross-reference.10. Research stub fill-in
I left [TK] markers for facts I need to verify. For each one, tell me: what
claim it supports, what kind of source would back it (primary data, vendor
docs, a named study), and where I'd look. Do NOT invent statistics — only
suggest research paths.This is the safe way to use AI for “research.” It points you at sources instead of inventing them — which matters, because hallucinated citations still show up in over 30% of research-context chatbot answers as of 2026.
11. Original-thinking layer
This article restates known wisdom. Add 1-2 paragraphs of original
thinking: an angle the audience hasn't seen, a counterintuitive framing, or
a synthesis no competing piece covers. If you genuinely can't add one,
reply "No original angle yet" instead of forcing it.The escape hatch is the point. A forced “original” angle is worse than none.
12. Bloat audit (run last)
Audit this expanded draft for bloat: (1) sentences that carry no
information, (2) adverbs that don't change meaning, (3) examples that repeat
a point already made, (4) sections that lost the thread. Output a cut list
with line references, not a rewrite.After expanding, this is the prompt that keeps you on the right side of Google’s density signal. Expect to cut 10-20% of what the earlier prompts added.
The fact-check pass is non-negotiable
Frontier model hallucination rates in 2026 run roughly 3.1% to 19.1% depending on the model and task — much better than the 15-45% of 2024, but nowhere near zero. The riskiest output is exactly what expansion prompts ask for: statistics and citations. A May 2026 Columbia University study in The Lancet found that by early 2026, 1 in 277 PubMed-indexed papers contained at least one fabricated reference, up from 1 in 2,828 in 2023 — a more than 12-fold rise tied to AI assistants.
The lesson for writers: treat every number and source the model adds as unverified until you check it. Prompt #2 and #10 are built to surface (not bury) the claims that need a human.
Common mistakes
- Vague audience. “Write for marketers” produces generic output. Name one specific reader.
- No tone anchor. Without 1-2 sample sentences in your voice, every variant comes back the same flavor.
- No constraints. Set word count, banned phrases, and a length cap, or the model defaults to long.
- Expanding a weak thesis. Fix the argument first; expansion only magnifies it.
- Trusting the first pass. The model lands on the safe middle. Push it.
- Shipping without a fact-check. AI is confident, not always correct.
How to push results further
- Give 1-2 real sentences in your voice. “Be friendly” is noise; a sample is a signal.
- Constrain ruthlessly — caps and banned phrases do more than any “make it good” instruction.
- Read the result aloud before publishing; padding is audible.
- Cut adverbs and adjectives that don’t change the meaning.
- Let AI do drafts 1 and 2; a human edits draft 3, and draft 3 is what ships.
- Anchor every prompt in one real person from your audience.
FAQ
- How long should the piece actually be? As long as it takes to fully answer the question and no longer. Google ranks how completely and quickly you satisfy intent, not word count.
- Which model is best for expanding a long draft? For drafts plus sources over ~5,000 words, use Claude Opus 4.7, Sonnet 4.6, or Gemini 3.1 Pro — all hold 1M tokens. ChatGPT Plus truncates older context unless you’re on the $200 Pro tier.
- Can AI write the whole thing? Use it for passes 1 and 2 (structure, evidence suggestions, examples). A human must own pass 3 — voice, judgment, and the fact-check.
- How do I stop the model from inventing statistics? Use prompts #2 and #10: tell it to name the source or flag the claim, and to suggest research paths rather than produce numbers. Then verify every figure.
- How often should I refresh an expanded piece? When the audience or the claims change, or quarterly for evergreen content carrying any pricing or version data.
- Can I reuse these prompts for other content types? Yes. Swap audience, goal, and voice; the structure (find gaps → fill with evidence → audit bloat) is content-agnostic.
Related
- Blog outline prompts
- Article rewrite prompts
- Blog introduction prompts
- Tone rewrite prompts
- Writing & Copywriting Prompts hub
- External: Google’s Helpful Content guidance
Tags: #Prompt #Writing #Long-form #Expansion