The 13 workflow files from the talk.

triage.md

# Triage every page

**You need:** 16 months of Search Console, with AI impressions.

## Instructions for the AI
Give each page in my Search Console export one decision: keep (page one, clicked), improve (page two), rebuild (in AI answers, rarely clicked), retire (past page two, not in AI) or merge (same query as another page). Table: page, decision, the numbers behind it.

## Read the answer
- Keep: refresh its evidence every year.
- Improve: add what only you have: your data, your tests, your people.
- Rebuild: AI answers use it but nobody clicks. Rebuild it around what only you can say.
- Retire or merge: redirect the weaker page to the stronger one.
- Check every number against your export before you act, and look at each retire and merge yourself first.

## Make it yours
Change the five decisions, or the rule in brackets after each one, to match how your site works. Keep one decision per page.

## How we tested it
A fresh Claude Sonnet session, 2 October 2026, on a five-page sample export (our teaching demo topic, commercial heat pumps): five pages, five decisions, each matching its rule. The page-two page came back as "Position 14.0 (page two), 40 clicks on 6,200 impressions, 0 AI impressions". We ran this whole file again the same day on a five-page sample export: the same five decisions.

## Use it
Paste this whole file into a new chat in any AI tool, add your export at the end, send.

## Your input
Export: [paste your export: page, top query, clicks, impressions, average position, AI impressions]

make-test.md

# The Make Test

**You need:** a brief's title, and what only you have on it.

## Instructions for the AI
For the brief below, check the top results, then answer in order, stopping at the first no: 1 Is what's there thin or badly sourced? 2 Is what only we have missing from all of them? 3 What must the piece answer? Quote what decides each. JSON.

## Read the answer
- No at question 1: the topic is already covered well. Do not write it.
- No at question 2: anyone could write it as well as you. Do not write it, or first get something only you have.
- An answer to question 3: write it, and give the writer that answer as the brief.
- Missing a content quota is fine. Filling it with weak pieces is not.

## Make it yours
Be honest in "Only we have". The model cannot know what only you have; it can only push back on what you tell it.

## How we tested it
A fresh Claude Sonnet session with web search, 2 October 2026, on Google's own example of commodity content, "7 Tips for First-Time Homebuyers", with "Only we have: nothing beyond general knowledge". It read the top nine results and quoted them. Question 1: yes, thin ("6 of the 8 tip articles among the top 9 results are short or cite nothing for their figures"). Question 2: no ("The brief lists no exclusive asset."). It stopped there. We ran this whole file again the same day: yes to question 1, no to question 2, stopped.

## Use it
Paste this whole file into a new chat with web search on, fill in the two lines at the end, send.

## Your input
Brief: [title]
Only we have: [our data, tests, people]

keyword-test.md

# The single-keyword test

**You need:** two or more sites in one niche, and each one's idea.

## Instructions for the AI
For the keyword below, write the angle each of my sites would take. Keep the facts the same on every site. Answer in JSON: site, angle, and the one site a reader could swap it with without noticing, or null.

## Read the answer
- A site named as swappable shares its angle with another site. Fix the idea, not the article.
- Act on the site that comes back in every run. One answer from a model is noise.

## Make it yours
- Pick a hard keyword: a term whose facts cannot change, so every site has to say the same true things.
- Give it each site's whole idea: its core idea, its instinct and what its reader leaves with. On one-line ideas alone it flags far more pairs.

## How we tested it
Three fresh Claude Sonnet sessions, 2 October 2026, on our own seven sites and a real keyword (both kept private). The flagged pair changed every run: D with F, then D with G, then A with F and D with G. Site D was in all three.

## Use it
Paste this whole file into a new chat in any AI tool, fill in the two lines at the end, send. Run it three times.

## Your input
Keyword: [keyword]
My sites: [each site's idea and what its reader leaves with]

claim-checker.md

# The claim checker, or Skeptic

**You need:** your claims, each with its source text.

## Instructions for the AI
Check each claim below against its source, never your memory. If you are unsure, the verdict is unsupported. Never fix a claim. Answer in JSON: claim id, verdict, and the reason, quoting the source.

## Read the answer
- Supported: the writer may use the claim.
- Unsupported: the claim is dead, or it becomes something to commission: the data to pull, the expert to call.
- If a quoted source line is not in your source, throw the answer away.

## Make it yours
Never show this chat the draft. It judges claims, not prose.

## How we tested it
In our teaching demo, August 2026, a longer version of this checker returned claim R8 as unsupported: "the cited product information sheet lists maximum current as 40 A, not the claimed 16.6 A". This short form has not been run on that claim. Run it on claims you already know the answer to before you rely on it.

## Use it
Paste this whole file into a new chat, never the one that wrote the draft, add your claims at the end, send.

## Your input
Claims: [your ledger rows, each with its source]

scissors-test.md

# The scissors test

**You need:** one section: a heading and the text under it.

## Instructions for the AI
Read this section as if the rest of the page were gone. Verdict SURVIVES if it still answers a question with its source and date, otherwise DIES. List each rule it breaks, quoting the words: answer first; heading answers; name the thing; date, place and source inside; one idea; define first; tables compare. Answer in JSON: verdict, and each broken rule with its quote.

## The seven rules
1. Answer first: the first sentence answers the question.
2. Heading answers: the heading states the answer.
3. Name the thing: never "it", use the name.
4. Date, place and source inside: all three in the fact's own sentence.
5. One idea: one idea per section.
6. Define first: define a term before you use it.
7. Tables compare: anything compared goes in a table.

## Read the answer
- DIES: fix it with the rewrite file, then run this test again.
- SURVIVES with broken rules: it stands alone, and the list is your polish.
- If a quote is not in your section, throw the answer away.

## How we tested it
Fresh Claude Sonnet sessions, 2 October 2026, on the three passages from the talk: DIES, DIES, SURVIVES. The verdict and the rule list are separate on purpose: our first version killed any section that broke any rule, including one that does stand alone. We ran this whole file again the same day on the same three passages: the same three verdicts.

## Use it
Paste this whole file into a new chat in any AI tool, add your section at the end, send. One section per chat.

## Your input
Section: [one heading and its text]

rewrite.md

# The rewrite

**You need:** a section that failed the scissors test.

## Instructions for the AI
Rewrite this section so it stands alone: answer first, the thing named, and the date, place and source in the sentence. Use only facts already in it. List any that are missing; never guess. Answer in JSON: text, facts used, missing.

## Read the answer
- A rewrite: check each fact against the original, then run the scissors test again.
- A refusal with a missing list: good. The missing list is your research brief.

## How we tested it
A fresh Claude Sonnet session, 2 October 2026, on passage 1 from the talk. It refused to rewrite: four facts it could use and seven missing (the thing being compared, the alternative it is cheaper than, the cost figures, the date or period, the place or market, the source, the table contents). We ran this whole file again the same day on passage 1: again no rewrite, and seven facts missing.

## Use it
Paste this whole file into a new chat in any AI tool, add the section at the end, send.

## Your input
Section: [one heading and its text]

citation-teardown.md

# The citation teardown

**You need:** a passage AI cited from a rival, and your section.

## Instructions for the AI
Passage A was cited in an AI answer. Passage B is mine, on the same question. Compare them rule by rule: answer first; heading answers; name the thing; date, place and source inside; one idea; define first; tables compare. For each, quote the deciding words and name the winner. JSON.

## Read the answer
- Rewrite your section where B lost, rule by rule. Check the same search again in two weeks.
- It tells you where your section loses, not why Google chose the other one. Nobody can see that.

## How we tested it
A fresh Claude Sonnet session, 2 October 2026. A was passage 3 from the talk; B was "As we covered above, this scheme offers generous support for businesses looking to upgrade." A won 4 rules, B none, one tie, and two rules did not apply. We ran this whole file again the same day on the same two passages: the same score.

## Use it
First find the cited passage: search your question and open the AI Overview's sources, then open the rival's link. The text after #:~:text= in the address is the passage it used. Then paste this whole file into a new chat in any AI tool, add both passages at the end, send.

## Your input
A: [cited passage]
B: [your section]

research-brief.md

# The research brief

**You need:** a topic, and an AI tool with web search on.

## Instructions for the AI
Research the topic below on the live web. Answer in JSON with two lists. confirmed: each fact with its source URL and the exact sentence that supports it. do_not_claim: things that sound true but you could not confirm, and why. Nothing goes in confirmed without a quote.

## Read the answer
- Give the writer both lists. Read do_not_claim first: it tells the writer what it must not say.
- Open a few source URLs and check the quotes are really there before you trust the brief.
- One brief can feed every language you publish in.

## How we tested it
A fresh Claude Sonnet session with web search, 2 October 2026, on "the grant a UK business can get for an air source heat pump in 2026" (our teaching demo topic): 39 confirmed facts, each with its URL and quote, and 17 things not to claim, each with the reason. One of them: do not claim that a business's heat pump installation is 0% VAT, because the relief covers residential and charitable buildings only (HMRC VAT Notice 708/6). We ran this whole file again the same day: 49 confirmed facts and 18 things not to claim, the VAT claim among them. The counts change from run to run; every confirmed fact should still carry its quote.

## Use it
Paste this whole file into a new chat with web search on, add your topic at the end, send.

## Your input
Topic: [topic]

claim-extraction.md

# Pull out every claim

**You need:** a draft, or any text you are about to publish.

## Instructions for the AI
List every factual claim in the text below, one per line, each rewritten so it makes sense on its own. Do not judge them, fix them or add anything. Answer in JSON: claim, and what kind of source could prove it.

## Read the answer
- Each claim is now a row in your evidence ledger, ready for the claim checker.
- A claim nobody could source is something to cut, or something to commission.

## How we tested it
A fresh Claude Sonnet session, 2 October 2026, given this whole file and passage 2 from the talk. Five claims came back, each standing alone, each with the kind of source that could prove it. The first: "The vast majority of searches now end without a click."

## Use it
Paste this whole file into a new chat in any AI tool, add the text at the end, send.

## Your input
Text: [paste text]

writer.md

# Write only what you can prove

**You need:** your checked ledger rows, and a section heading.

## Instructions for the AI
Write the section below using only these ledger rows. After each sentence, put its row id in brackets. If the section needs a fact no row has, write [MISSING: what] instead. Never use your own knowledge.

## Read the answer
- Every sentence should carry a row id. A sentence without one came from the model's memory: delete it.
- Each [MISSING: ...] line is research to do, not a gap to fill by hand.

## How we tested it
A fresh Claude Sonnet session, 2 October 2026, given this whole file, three real rows from our teaching demo's ledger and a section on what an air source heat pump costs a business. Every sentence carried its row id, for example "The grant is for individual systems up to 45kWth. [R1]". Where no row had the fact, it wrote four [MISSING: ...] lines instead of guessing, among them the installed price and the yearly running cost.

## Use it
Paste this whole file into a new chat with web search off, add the rows and the heading at the end, send.

## Your input
Rows: [your ledger rows]
Section: [heading]

link-check.md

# Catch invented links

**You need:** your sitemap's URLs as one list, and a draft.

## One line for your writing prompt
Link only to URLs on this list. None fits? Write [NO LINK].

## Instructions for the AI
Here is my list of real URLs and a draft. List every link in the draft that is not on the list, exactly as written, and the sentence it sits in. Answer in JSON.

## Read the answer
- Every link it returns fails: fix it or remove it.
- Rebuild the list whenever pages go live or come down.

## How we tested it
A fresh Claude Sonnet session, 2 October 2026, on a test draft with four example.com links, one of them made up. It returned only the made-up one, with its sentence. We ran this whole file again the same day on a new test draft, four links with one made up: it returned only that one.

## Use it
Add the line above to your writing prompt. Before anything publishes, paste this whole file into a new chat, add the list and the draft at the end, send.

## Your input
URLs: [your sitemap list]
Draft: [paste]

query-topics.md

# Group your queries into topics

**You need:** up to 500 queries from Search Console.

## Instructions for the AI
Group these search queries into 8 to 12 topics. For each topic: a short name and one regex that matches its queries. Then list the queries that fit no topic. Answer in JSON.

## Read the answer
- Freeze the regexes. Use each one in Search Console's query filter (custom regex) and compare the same topics every month.
- New queries that match no topic are your early warning.

## How we tested it
A fresh Claude Sonnet session, 2 October 2026, given this whole file and 40 sample queries about commercial heat pumps (our teaching demo topic): eleven topics, every query placed. One of them: "Running costs", regex running cost|cost to run|electricity cost|kwh|energy bill.

## Use it
Paste this whole file into a new chat in any AI tool, add your queries at the end, send. Do it once, then keep the result.

## Your input
Queries: [up to 500 queries from Search Console]

buyer-questions.md

# Your buyers' questions

**You need:** what you sell, and who buys it.

## Instructions for the AI
List 20 questions a buyer of the product or service below asks an AI tool before choosing a provider, written the way they would type them. No brand names. Group them: comparing options, cost, trust, how it works.

## Read the answer
- These 20 are your prompt panel. Ask each one five times in each AI tool you care about.
- Count the runs that mention you. Report that share every month, never a rank.

## How we tested it
A fresh Claude Sonnet session, 2 October 2026, for "commercial air source heat pumps, a facilities manager for a UK office building" (our teaching demo topic). Twenty questions came back in the four groups, for example "will an installer guarantee the energy savings they quote or is it just sales talk". We then asked one of them five times, each in a fresh session with web search: 25 sites were cited, and only 3 in every run. We ran this whole file again the same day: twenty questions in the four groups.

## Use it
Paste this whole file into a new chat in any AI tool, fill in the line at the end, send.

## Your input
Product or service: [product or service, and who buys it]