Oscar Craven
Menu
Who’s looking?

The page reorders for you. Links keep your choice.

Work

Before and after with behavioural science

An AI skill that rewrites a page, and names the study behind every change.

Built and tested, October 2026

Two minutes? Read the five lines, then flip the first rewrite.

See the three rewrites Download CV

The short version

What it is
An AI skill that reads a web page or app flow and rewrites it using tested behavioural science. Every change shows the study it rests on.
Who for
Founders and product teams who want more sign-ups, upgrades or repeat orders without tricking anyone.
What I did
I read 57 newsletters by two behavioural scientists. I turned them into a 40-rule library with sources. Then I built the skill and tested it on three mock sites.

Emails from Richard Shotton and Patrick Fagan, February 2025 to October 2026. Library as of 11 October 2026.

What happened
With the skill, the AI passed 30 of 30 checks. Without it, 13 of 30.

Three mock briefs, ten checks each, graded by an AI model on 11 October 2026. No human grading yet.

What I’d change
Test on real traffic next. Every number in the library is a starting bet for an A/B test, not a promise.

Open the before pageOpen the after page

Before
Ledgerly pricing page, before Top of Ledgerly’s mock pricing page, before the rewrite: Basic at £15 a month beside Pro at £348 a year.
After
Ledgerly pricing page, after Top of Ledgerly’s mock pricing page, after the rewrite: a new headline and three plans. Pro sits in the middle, marked “Recommended for agencies with 4+ clients”.

Ledgerly’s pricing page. Mock site built for the test. Names, figures and reviews are fictional.

Three rewrites

Three mock sites, each with a founder’s brief. Press After to see what the skill changed. The numbers follow the skill’s own ranked list.

Ledgerly: more people choosing Pro

A pricing page for a small UK app. The brief: “Almost everyone picks Basic or bounces. I want more people choosing Pro and fewer leaving.”

Before is the page as briefed. After is the skill’s rewrite.

Before
Ledgerly pricing page, before Ledgerly’s mock pricing page, before the rewrite. Two plans: Basic at £15 a month and Pro at £348 a year. Below them: “Trusted by over 1,000 agencies.”
After
Ledgerly pricing page, after Ledgerly’s mock pricing page, after the rewrite. Three plans: Starter at £15, Pro at £29 and Agency at £49 a month. Pro sits in the middle, marked “Recommended for agencies with 4+ clients”. A “Which plan?” line and a cancel line sit below.

Open the before pageOpen the after page

Mock site built for the test. Names, figures and reviews are fictional.

  1. Change 1: Show Pro per month and per day

    Before
    “£348/year”, beside Basic at “£15/month”
    After
    “£29 /month, billed annually”, “£348 a year, that’s 95p a day” and “+£14 a month more than Starter”

    Why: a small unit feels cheaper, and the total stays in view. Hardisty, University of British Columbia, 2019, 253 people, New York Times subscriptions. At full price, 22% picked the dearer plan. Shown as the extra it costs, 47% did.

  2. Change 2: Add a dearer plan, so Pro sits in the middle

    Before
    Two plans. Pro is the dearest on the page.
    After
    “Agency”, at “£49 /month, billed annually”, to the right of Pro

    Why: people avoid the cheapest and the dearest, so the middle gains. Astroten with Transmission and NewtonX, 2025, 500 B2B decision makers. Adding a dearer tier took Pro from 28% to 59%.

  3. Change 3: Recommend one plan, and say who it is for

    Before
    No recommendation.
    After
    “Recommended for agencies with 4+ clients” and “Which plan? More than 3 clients: Pro.”

    Why: people copy others like them, and a clear answer helps them choose. Behavioural Insights Team with HMRC, 2011, letters to 140,000 taxpayers. Paying on time rose from 67.5% to 72.5% with a general line, and to 83% with a tailored one.

  4. Change 4: Make the crowd precise

    Before
    “Trusted by over 1,000 agencies.”
    After
    “1,214 UK agencies reconciled £38.6m of client invoices with Ledgerly last month.”

    Why: exact numbers are easier to believe. Astroten with Transmission and NewtonX, 2025, 500 B2B buyers. “40% faster” was credible to 38%. “39.6% faster” was credible to 50%.

  5. Change 5: Show what Starter lacks

    Before
    Basic lists only what it has.
    After
    Two Pro features, struck through. Then: “Starter checks once a month, so a payment missed on the 2nd shows up on the 30th.”

    Why: a loss weighs more than the same gain. Gonzales, Aronson and Costanzo, 1988, a US energy firm. Told what they would save, 39.1% took the offer. Told what they would lose, 60.9% did.

Fernhill Roasters: more first boxes, more subscriptions

A coffee subscription product page. The brief: “more first-time visitors to add the 6-bag box to basket and more of them to pick the subscription”.

Before is the page as briefed. After is the skill’s rewrite.

Before
Fernhill product page, before Fernhill Roasters’ mock product page, before the rewrite. A one-off purchase and a subscription, both at £47.99. An “OUT OF STOCK” label sits over a grey Add to basket button. A £2.49 supermarket coffee leads the comparison strip.
After
Fernhill product page, after Fernhill Roasters’ mock product page, after the rewrite. The price adds £8.00 a bag and about 48p a cup. The subscription comes first, marked “Most popular with first-time buyers”. “Sold out. Next roast Thursday 22 October.” sits above two live buttons.

Open the before pageOpen the after page

Mock site built for the test. Names, figures and reviews are fictional.

  1. Change 1: Give the subscription a real advantage

    Before
    One-off first, then “£47.99 every 6 weeks (free shipping, free Fernhill mug)”
    After
    “Every 6 weeks” first, marked “Most popular with first-time buyers”. The one-off shows “No mug”.

    Why: people copy buyers like them. Behavioural Insights Team with HMRC, 2011, 140,000 letters. A generic line took payment from 67.5% to 72.5%. A tailored one took it to 83%.

  2. Change 2: Put a price on the freebies

    Before
    “free Fernhill mug” and “Free shipping on orders over £40”
    After
    “Fernhill mug in your first box (sells for £12)” and “Free UK shipping on this box (saves £4.95).”

    Why: free on its own reads as worthless. Raghubir, NYU, 2004. A free bracelet was rated 55% less valuable when bundled. Stating the price cut that penalty.

  3. Change 3: Do the maths: per bag and per cup

    Before
    “£47.99”
    After
    “£47.99 · £8.00 a bag · about 48p a cup”

    Why: busy people never divide. Astroten with Flicker, 2022, 282 people. Adding “$1.58 a bottle” to a 12-pack made the deal look more than twice as good.

  4. Change 4: Drop the supermarket coffee from the comparison

    Before
    “Supermarket Own Brand Ground Coffee 227g”, at “£2.49”
    After
    Three Fernhill coffees, each priced by the bag, such as “£11.95 (£11.95 a bag)”

    Why: a price looks good or bad next to its neighbours. Astroten with Flicker, 2022, 404 shoppers. Ben & Jerry’s at $3.99 was good value to 41% beside $4.99 Halo Top. Beside $1.99 own-label, it was 27%.

  5. Change 6: A sold-out state that still sells

    Before
    “OUT OF STOCK (back soon)” over a grey “Add to basket” button
    After
    “Sold out. Next roast Thursday 22 October.” with “Reserve my first box” and “Email me when it’s back”

    Why: “sold out” suggests other people wanted it. Peterson, University of Texas, 2019, 1,117 people. It caused 11% less disappointment than “out-of-stock”, and 15% less than “unavailable”.

Stride: more walkers still there on day 7

A walking-habit app’s onboarding and first week of notifications. The brief: “Day-1 to day-7 retention is 38% and I think the onboarding and notifications are the problem.”

Before is the flow as briefed. After is the skill’s rewrite.

Before
Stride onboarding, before Stride, a mock walking app, before the rewrite. Five onboarding screens, including “Stay motivated” and a leaderboard where you rank 48,212 of 48,212. Below: a week of notifications at fixed times.
After
Stride onboarding, after Stride, a mock walking app, after the rewrite. Five onboarding screens, including “When will you walk tomorrow?” with time and place choices. Below: a week of notifications sent ten minutes before the planned walk.

Open the before pageOpen the after page

Mock site built for the test. Names, figures and reviews are fictional.

  1. Change 2: Plan the walk: a time and a place

    Before
    “Stay motivated” and a “Daily reminders” switch
    After
    “When will you walk tomorrow?”, then “Set my walk: 07:45, to the station”

    Why: a plan tied to a moment gets done. Milne, Orbell and Sheeran, 2002, University of Bath, 248 people. 35% exercised with no prompt and 38% with motivational information. With a when-and-where plan, 91% did.

  2. Change 3: Send the nudge ten minutes before the walk

    Before
    Fixed times, such as 09:00: “Don’t forget to walk today!”
    After
    One push a day at 07:35, such as “Walk two. To the station, 07:45.”

    Why: the nudge lands at the moment the plan names. Same study: Milne, Orbell and Sheeran, 2002.

  3. Change 4: No scolding after a missed goal

    Before
    “You missed your goal today. Try again tomorrow.”, then a Stride Plus offer
    After
    “Not today.” and “Tomorrow’s walk is set: 07:45, to the station. Keep it, move it, or shrink the goal?”

    Why: mood colours judgement. Shotton with News UK and Carat, 2023, 821 people. In a good or very good mood, people were 27% more likely to rate an offer good value.

  4. Change 5: Offer Plus once, after a win

    Before
    “Missed your goal? Stride Plus can help”
    After
    After the third goal: “7,000 steps, 4 days running. Want to see the pattern?” and “£4.99 a month, that’s 16p a day”

    Why: an offer after a win feels better value, from the same 821-person study. The daily price draws on Shotton’s 500-person test. A daily frame was about 5 times more likely to be rated a great deal.

Results

Each brief ran once with the skill and once without it.

Each brief had ten checks. Examples: “cites the study and its figures”, “no invented statistics”, “names a primary and a guardrail metric”. Graded by an AI model from the outputs on 11 October 2026. Not yet reviewed by a person.
BriefWith the skillWithout the skill
Ledgerly pricing page10 of 102 of 10
Fernhill product page10 of 104 of 10
Stride onboarding10 of 107 of 10
Total30 of 3013 of 30

Read this before the score

  • Graded by a model, not a person yet.
  • Most of the library’s evidence is a survey of a few hundred people imagining a purchase, not a field trial.
  • The skill costs about 1.6 times the tokens and twice the time of a plain answer.

Mean of three runs each, 11 October 2026: 99,820 tokens with the skill, 60,816 without; 431 seconds against 203.

How it was built

Five steps, from reading to a tested skill.

  1. 1. Read 57 emailsRichard Shotton’s AstroHacks, December 2025 to October 2026. Patrick Fagan’s newsletters, February 2025 to June 2026.
  2. 2. Pull out each principleWith its study, its sample and its effect size.
  3. 3. Write the skillA 40-rule library, a surface checklist, a 14-row rubric with an ethics gate, and a report template.
  4. 4. Write three mock briefsEach one has ten checks a good answer must pass.
  5. 5. Run, grade, fixRun each brief with and without the skill. Grade it, fix the skill, log it.

Two rules the skill lives by

Every number is a starting bet, not a promise

The library quotes each figure as its study reports it. Most come from small surveys. So each one is the first guess for an A/B test.

A nudge that only works because the user misunderstands has failed

Every change must pass an ethics gate. A change that fails is removed, not softened. The floor is UK consumer law: CMA guidance, the ASA codes and the Consumer Protection from Unfair Trading Regulations 2008.

The library

Each rule says what to change and the study behind it.

Five of 40. Each row names its study so a product team can check it. Figures as each study reports them; library as of October 2026.
PrincipleWhat it saysEvidence
Per-day pricePeople judge the size of the number in front of them.Shotton, 500 people. A daily price was almost five times more likely to be called a great deal than a yearly one.
Sold out, not unavailable“Sold out” blames popularity. “Unavailable” blames the shop.Peterson, 2019, 1,117 people. “Sold out” caused 15% less disappointment than “unavailable”.
A tailored crowdPeople copy others, most of all people like them.HMRC letters with the Behavioural Insights Team, 2011, 140,000 taxpayers. “9 out of 10 people in your town pay on time” lifted payment from 67.5% to 83%.
Say the flaw firstAdmitting a weakness makes the rest easier to believe.Fennis, 2014. When a firm admitted a side effect itself, 59% took its free health check. When a third party revealed it, 13% did.
When and whereA plan tied to a time and a place gets done.Milne, Orbell and Sheeran, 2002, 248 people. Asking when and where they would exercise took weekly exercise from 35% to 91%.

What is next

The product. Paste in a URL and get the report in minutes, with the ethics gate built in. Done by a person, a review like this is days of work.

The next test. Real traffic. Each change runs as an A/B test, with the library’s figure as the starting bet. The result then shows whether the study holds for that page.

Want this run on your page?

Tell me what the page is for. I read every message.

Send me a message See how I work