How to use this library
Start with one page and one reader. Do not paste all of these prompts into a single conversation and ask for a finished article. That usually produces a long document with confident transitions, repeated advice, and no clear record of where its claims came from.
Instead, save the output from each stage. Keep the query brief, evidence ledger, outline, draft, challenge report, and refresh notes together. Use a model that can handle the amount of context you provide, but do not confuse a larger context window with better research. Give the model source URLs, excerpts, test notes, and your own constraints. Ask it to mark uncertainty rather than fill gaps.
Decide
Define the reader's decision and the page's job before asking for copy.
Investigate
Build an evidence ledger and identify what you know, need to test, or should leave out.
Draft
Turn the evidence into a page with a direct answer, useful examples, and a next step.
Challenge
Try to break the page: unsupported claims, missing conditions, vague headings, and weak advice.
Maintain
Record what will become stale and define the next refresh trigger before publication.
Stage 1: decide what the page must do
A keyword is not a page brief. "Best AI tools for recruiters" could mean a buyer wants a shortlist, a recruiter wants help with one task, a manager wants a procurement framework, or a candidate wants tools to prepare for an interview. The prompt below forces those jobs apart before you draft.
Prompt 1: turn a query into a reader decision
Use it for: intent mapping, scope, and choosing the correct page type
You are the editorial strategist for a people-first website.
Topic or query: [QUERY]
Audience: [SPECIFIC READER]
Business context: [WHAT WE OFFER, IF RELEVANT]
Known supporting pages: [URLS]
Return:
1. The most likely reader situation in one paragraph.
2. Three plausible intents, ranked by usefulness and commercial relevance.
3. The decision or task the reader wants to complete after reading.
4. The page type that best serves each intent: guide, comparison, review, template, calculator, directory, or troubleshooting page.
5. What belongs on this page and what should become a separate supporting page.
6. Five follow-up questions an AI-search user may ask next.
7. One sentence describing what would make this page genuinely better than a generic answer.
Do not draft the article. Flag assumptions and ask for missing context.Human review: Check whether the proposed reader sounds like a real person with a real task. Delete intents that are only keyword variations. Approve one primary job before moving on.
Prompt 2: define a useful promise
Use it for: writing a precise title, opening answer, and success criterion
Using the approved reader job below, write three versions of a page promise.
Reader job: [APPROVED JOB]
Page type: [TYPE]
Audience expertise: [BEGINNER, PRACTITIONER, OR EXPERT]
Geography or market: [LOCATION, IF RELEVANT]
For each version provide:
- A clear title without hype.
- A one-sentence answer to the core question.
- The evidence or experience the page must contain to earn that promise.
- A measurable "reader can now..." outcome.
Do not promise rankings, citations, guaranteed ROI, or results we cannot prove. Prefer a narrower promise that we can actually fulfill.Human review: Choose the promise that describes the work the page will do, not the traffic you hope it will receive. If the evidence requirement is impossible, narrow the promise.
Stage 2: build the evidence before the prose
The evidence ledger is the most important artifact in this library. It prevents a model from turning a plausible statement into a published fact. Give each claim a type, scope, source, date, confidence, and owner. If you cannot fill those fields, the claim is not ready to carry the page.
Prompt 3: create an evidence ledger
Use it for: separating fact, observation, recommendation, and open question
Create an evidence ledger for this page brief.
Page brief: [BRIEF]
Research notes and URLs: [PASTE NOTES AND LINKS]
Internal experience or test data: [PASTE TEST NOTES]
Return a table with these columns:
- Claim in plain language
- Claim type: fact, observation, recommendation, or interpretation
- Exact scope: who, where, when, and under what conditions
- Best supporting source or test note
- Source date or version
- Confidence: high, medium, or unresolved
- What would change the claim
- Editorial action: publish, qualify, test, source, or remove
Do not invent citations. If a URL does not support the claim, mark it as a mismatch. Keep unresolved claims out of the draft until a human decides what to do.Human review: Open every important source. A source that mentions the same topic is not necessarily evidence for your exact sentence. Replace weak secondary summaries with primary material when possible.
Prompt 4: find the information gain
Use it for: turning competitor research into original reader value
Compare the following pages only as research inputs, not copy sources.
Target query: [QUERY]
Competing URLs: [URLS OR NOTES]
Our evidence ledger: [LEDGER]
Our experience, test, data, or access: [ORIGINAL MATERIAL]
Return five sections:
1. Questions every page answers.
2. Important questions most pages answer vaguely or skip.
3. Claims that look repeated, unsupported, or stale.
4. Original contributions we can responsibly make from our evidence.
5. A prioritized list of page additions, each tied to a reader decision.
Do not recommend copying headings, phrasing, examples, or structure. If we have no original material for a proposed addition, label it as a research gap.Human review: Require a concrete contribution: a test, local example, comparison criterion, source synthesis, worked example, or decision aid. "More detail" is not information gain by itself.
Stage 3: draft for a person who may quote you
AI systems may extract a sentence from a page, but the person who clicks through still needs context. Write answer-first sections that stand alone without becoming a stack of disconnected snippets. Each section should resolve a question, show the condition or tradeoff, and tell the reader what to do next.
Prompt 5: make a research-shaped outline
Use it for: creating sections that answer questions and progress a decision
Build an outline for this approved brief.
Brief: [BRIEF]
Evidence ledger: [LEDGER]
Information-gain opportunities: [LIST]
Related pages available: [URLS]
For each section provide:
- A specific H2 phrased as the reader's question or task.
- The direct answer in 1 to 3 sentences.
- The evidence or example that belongs there.
- The limitation or tradeoff the reader should know.
- The next action or internal link, if one genuinely helps.
Start with the core answer, then explain the decision criteria, show examples or comparisons, handle edge cases, and finish with implementation and maintenance. Do not use headings such as "Why it matters," "Best practices," or "Conclusion" unless the section has a specific job.Human review: Read the outline without the topic label. If the sequence does not make sense to a busy reader, fix it before asking for prose. Remove sections that repeat the same answer.
Prompt 6: draft one section with source boundaries
Use it for: producing clear copy without laundering uncertainty
Draft only this section, not the full article.
Section question: [H2]
Reader and decision: [CONTEXT]
Approved evidence: [RELEVANT LEDGER ROWS]
Sources: [URLS AND EXCERPTS]
Required example or test: [DETAILS]
Rules:
- Answer the section question in the first sentence.
- Distinguish sourced fact from our observation and recommendation.
- Use the exact scope and date in the evidence ledger.
- Name a condition, tradeoff, or limitation where one matters.
- Do not add numbers, features, quotes, or citations that are not in the supplied evidence.
- Use plain language and paragraphs that can be understood when quoted with the heading.
- End with a practical next step only if it follows from the evidence.
After the draft, list any sentence that still needs human confirmation.Human review: Verify every sentence against the ledger. Keep the model's uncertainty list. A polished paragraph is not approved merely because it sounds authoritative.
Stage 4: challenge the page before publishing
Most publishing workflows ask an AI tool to make a draft better. I get more value by asking it to make the draft uncomfortable. The challenge pass should find the sentence that overreaches, the heading that says nothing, the internal link that leads nowhere, and the recommendation that lacks a criterion.
Prompt 7: run a hostile but fair quality review
Use it for: finding unsupported claims, repetition, and reader disappointment
Review this draft as a demanding reader and a careful editor.
Target query and audience: [QUERY, AUDIENCE]
Draft: [PASTE DRAFT]
Evidence ledger: [LEDGER]
Sources: [URLS]
Return a table with:
- Exact sentence or section
- Problem type: unsupported claim, scope error, stale detail, repetition, vague heading, weak example, missing condition, poor link, or unnecessary filler
- Why it could mislead or disappoint the reader
- Severity: blocking, important, or optional
- Precise edit needed
Then answer:
1. What would a reader still be unable to do after reading this?
2. What is the page's strongest original contribution?
3. Which paragraph is most likely to be quoted out of context, and what qualification does it need?
4. Which section should be removed if it cannot be improved with evidence?
Do not rewrite the whole page. Diagnose it first.Human review: Accept only findings tied to the audience and evidence. Apply the blocking fixes yourself, then rerun the review on the changed sections.
Prompt 8: check AI-search extractability without writing for robots
Use it for: testing whether useful answers are easy to locate and understand
Audit this page for a person arriving from an AI-generated answer.
Page: [PASTE DRAFT OR URL]
Primary question: [QUESTION]
Sources and evidence ledger: [MATERIAL]
Check:
- Can the first screen answer the primary question?
- Does every important section have one clear question and answer?
- Can a quoted paragraph be understood with its condition and date?
- Are recommendations explained by criteria rather than presented as universal truth?
- Are tables, lists, and examples decision-useful rather than decorative?
- Does the page offer a better next step than "read more"?
- Would a visitor trust the page enough to inspect the sources?
Return the three highest-value edits. Do not add keyword repetition, fake FAQs, or claims about how a platform ranks pages.Human review: This is a usability test, not a request to make every paragraph short. Preserve nuance when it changes the decision.
Stage 5: localize, link, and maintain the winner
A page that performs in one market should not be translated mechanically into another. The reader may use different tools, payment methods, regulations, examples, vocabulary, and trust signals. Likewise, an internal link should be chosen because it solves the next problem, not because the site wants another URL in the paragraph.
Prompt 9: design a standalone local version
Use it for: localization that reflects behavior rather than word substitution
Create a standalone content brief for [TARGET LANGUAGE AND COUNTRY] based on this English topic.
English page or idea: [URL OR SUMMARY]
Local audience: [ROLE, EXPERIENCE, AND CONTEXT]
Known local constraints: [TOOLS, PRICING, PAYMENTS, REGULATION, SCHOOL OR WORK NORMS]
Local source material: [URLS, SEARCH QUERIES, COMMUNITY LANGUAGE]
Return:
- The local searcher's likely situation and vocabulary.
- What should be kept, changed, removed, or added.
- Local examples and decision criteria.
- A native-language outline with a different opening and section order where appropriate.
- Claims that require local verification.
- Five questions a local reader may ask next.
Do not translate the English page sentence by sentence. If local evidence is missing, mark the gap instead of inventing cultural detail.Human review: Ask a fluent subject-aware editor to check tone, terminology, search behavior, and whether the page would be useful to someone who never reads English.
Prompt 10: build an internal-link path
Use it for: connecting a page to useful hubs and next actions
Map internal links for this finished page.
Page purpose: [PURPOSE]
Full draft: [DRAFT]
Available site URLs and one-line descriptions: [URL LIST]
Return a table with:
- Source section
- Destination URL
- Descriptive anchor text
- The reader question it answers next
- Whether the link is essential, helpful, or should be omitted
Recommend no more than [NUMBER] links. Do not recommend a destination if its page does not fulfill the promised next step. Identify one missing supporting page only when the gap is real and worth building.Human review: Click every proposed link. Replace generic anchors and remove links that interrupt the reader's decision.
Prompt 11: create the refresh record
Use it for: keeping a strong page accurate after publication
Create a maintenance record for this page.
Page: [URL]
Current evidence ledger: [LEDGER]
Published claims: [LIST OR DRAFT]
Current date: [DATE]
Return:
- Volatile claims and their review interval.
- Source URLs to recheck.
- Product, pricing, policy, statistic, and platform changes that should trigger an immediate review.
- The metric or user feedback that would justify a substantive rewrite.
- Sections that should not be changed without a new test or source.
- A short change log template.
Do not recommend changing a page simply to alter its date. A refresh should improve accuracy, usefulness, or clarity.Human review: Assign an owner and a real date. A maintenance note that nobody can act on is documentation theater.
The final publication gate
Before publication, I want a person to be able to answer yes to the checks below. The model can help collect evidence, but the editor owns the decision. This is also the fastest way to catch a page that has been expanded to 2,500 words without becoming more useful.
The opening answers the reader's actual question instead of explaining the content strategy.
Important claims are linked to sources that support their exact scope and date.
Facts, observations from testing, recommendations, and interpretation are clearly separated.
The page contains information gain: a useful example, test, comparison, method, or decision aid.
Headings describe the question or task the section resolves, not a generic content label.
Internal links help the reader take the next step and lead to pages that really deliver it.
Volatile information has an owner, a review date, or a visible limitation.
A person with subject knowledge has reviewed the final claims and recommendations.
A useful rule
If the model can complete the page without your research notes, judgment, testing, or source decisions, the prompt is probably asking it to produce generic content. Add the missing human input or change the page idea.
What this process can and cannot do
This process can help a small editorial team work through complex topics consistently. It can expose intent gaps, organize sources, make a page easier to audit, create better briefs, and turn one strong piece of research into useful supporting pages. It can also make localization and maintenance less ad hoc.
It cannot manufacture first-hand experience, make an unsupported claim true, guarantee an AI citation, or make a thin page valuable by adding headings. Search platforms and readers respond to the quality of the underlying work. Prompts are a set of questions and controls around that work.
I would start with Prompts 1, 3, 5, and 7 on one existing page. If that small loop exposes weak evidence or a missing reader outcome, fix the page idea before producing more content. Scale the workflow only after the first page becomes more useful to a real person.
Sources and boundaries
These sources support the platform and editorial guidance in this library. Google says its AI search features use the same foundational Search practices, while its people-first and generative-AI guidance emphasize helpfulness, originality, accuracy, and human responsibility. Bing and OpenAI provide useful reporting and crawler context, but none publishes a prompt that guarantees a citation.
Frequently asked questions
Can these prompts make a page rank in Google or appear in an AI answer?
No prompt can guarantee rankings, citations, or traffic. These prompts improve the editorial process by making intent, evidence, usefulness, internal links, and maintenance easier to inspect. The page still needs a real audience, accessible publishing, accurate information, and time to earn visibility.
Should I paste a competitor page into an AI tool?
Use competitor pages as a source of questions, formats, and missing decision information, not as text to imitate. Ask what the competing page fails to answer and build from your own research, testing, experience, and sources. Never ask a model to disguise copied language as original work.
How long should an AI-search page be?
Long enough to solve the reader's job and no longer. A useful page may need 2,500 words when it includes comparisons, examples, evidence, and a workflow, but word count alone is not a quality signal. Remove repetition and keep the most important answer easy to find.
Can AI write the final article?
It can help with research organization, outlines, transformations, and first drafts. A human should still choose the angle, verify claims, add original knowledge, approve recommendations, check sensitive details, and take responsibility for the published result.
What is the best prompt in this library?
The evidence ledger prompt is the best place to start. It forces you to separate facts, observations, recommendations, and open questions before drafting. Most weak AI-search pages fail because the writer starts polishing prose before knowing which claims can be supported.