Why trust us

How we test and train AI workflows, what each badge means, and the evidence product by product: the numbers, the failures and the limits.

We test every workflow before we sell it, and we train the ones that matter most. Below: how, what each badge means, and the evidence product by product.

The loop

Real task → Baseline → Test → Inspect failures → Improve → Retest → Benchmark

The seven steps
  1. Real task. One job, real inputs, and a written idea of a good result before anything runs.
  2. Baseline. We ask a plain AI model to do the job with one line, and score it. That is the bar.
  3. Test. We run our workflow on the same cases and score every output against the same standard.
  4. Inspect failures. We read what went wrong and look for patterns, not one-off slips.
  5. Improve. We change a few things at a time.
  6. Retest. A change stays only if the score goes up on cases that did not shape it.
  7. Benchmark. At the end we run cases the workflow never saw, once, and publish the result.

Every product goes through steps 1 to 5 before release: at least one normal and one deliberately hard input, read against the source, fixed and re-run. Our flagships go through all seven.

The badges

BENCHMARKED. Measured against a plain AI request on cases it was not tuned on. The numbers are below.

OPTIMISED. Improved step by step against a set of test cases, with every score on file.

TESTED. Run on at least three real examples and scored against a written standard, with the record on file.

No badge. Checked on real input before release, but not yet on enough cases for a badge.

A product shows the strongest badge it has earned. A pack shows the lowest badge of what is inside.

What we found when we checked ourselves

  • One line is not enough. For product descriptions, a one-line request got 1 of 20 fully right on a strong model; our workflow got 20 of 20.
  • Hard inputs catch what easy ones miss. We re-ran 69 workflows on deliberately hard inputs. 16 had a defect a single run had missed, most often something marked "unsure" stated as fact. All 16 were fixed and re-run. Read it
  • Sensible advice can make things worse. In our trained browsing workflow, two of three sets of edits read well and made results worse. Only a score on unseen tasks showed it. Read it
  • Sometimes we lose. For late-payment emails, a one-line request beat our workflow on our checks (18 of 20 against 13 of 20). We are fixing it. The product does not change until the fix wins on cases it was not written from.

Evidence by product

Open a product for the cases, the result and the limits.

Browse the web with an AI agent in fewer steps BENCHMARKED badge. Measured against a plain AI request on cases it was not tuned on. The numbers are published.

A browsing helper for AI agents.

A browsing helper for AI agents.

Reply to reviews without promising what you can't give BENCHMARKED badge. Measured against a plain AI request on cases it was not tuned on. The numbers are published.

Reply to reviews without promising anything.

Reply to reviews without promising anything.

Rewrite thin product descriptions without made-up claims BENCHMARKED badge. Measured against a plain AI request on cases it was not tuned on. The numbers are published.

Rewrite thin descriptions, ready to paste.

Rewrite thin descriptions, ready to paste.

Turn call notes into a screen note (recruiters) BENCHMARKED badge. Measured against a plain AI request on cases it was not tuned on. The numbers are published.

Turns screening-call notes into a screen note and a client write-up.

Turns screening-call notes into a screen note and a client write-up.

Browse the web with a local Qwen model, fewer wrong turns OPTIMISED badge. Improved step by step against a set of test cases, with every score on file.

The same browsing helper, for a local Qwen model.

The same browsing helper, for a local Qwen model.
14 more products, each tested on real examples

Answer a chargeback: the right letter for each reason TESTED badge. Run on at least three real examples and scored against a written standard, with the record on file.

Answer a Shopify chargeback with the right letter for the reason.

Answer a Shopify chargeback with the right letter for the reason.

Gather the evidence for a chargeback TESTED badge. Run on at least three real examples and scored against a written standard, with the record on file.

Know what evidence to gather when a customer disputes a payment.

Know what evidence to gather when a customer disputes a payment.

See how AI shopping assistants read your store TESTED badge. Run on at least three real examples and scored against a written standard, with the record on file.

Score how AI shopping assistants read your Shopify store.

Score how AI shopping assistants read your Shopify store.

Quick check: can AI shopping assistants read your store? TESTED badge. Run on at least three real examples and scored against a written standard, with the record on file.

See how AI shopping assistants read your Shopify store, in six checks.

See how AI shopping assistants read your Shopify store, in six checks.

Plan your Black Friday sale on your own prices TESTED badge. Run on at least three real examples and scored against a written standard, with the record on file.

Plan your Black Friday and Cyber Monday sale on your own prices.

Plan your Black Friday and Cyber Monday sale on your own prices.

Handle the weekly jobs in a Shopify store: six workflows TESTED badge. Run on at least three real examples and scored against a written standard, with the record on file.

Handle weekly Shopify jobs: copy, SEO, support, disputes.

Handle weekly Shopify jobs: copy, SEO, support, disputes.

Clean up messy product titles TESTED badge. Run on at least three real examples and scored against a written standard, with the record on file.

Give every product in your Shopify store the same tidy title pattern.

Give every product in your Shopify store the same tidy title pattern.

Write image descriptions (alt text) for your products TESTED badge. Run on at least three real examples and scored against a written standard, with the record on file.

Write alt text for every product image in your Shopify store.

Write alt text for every product image in your Shopify store.

Check your store policies for gaps and mixed messages TESTED badge. Run on at least three real examples and scored against a written standard, with the record on file.

Find gaps in your refund and shipping policies.

Find gaps in your refund and shipping policies.

Rewrite your holiday let listing TESTED badge. Run on at least three real examples and scored against a written standard, with the record on file.

Rewrite your holiday-let listing title and description.

Rewrite your holiday-let listing title and description.

All 6 host tools TESTED badge. Run on at least three real examples and scored against a written standard, with the record on file.

Run your holiday-let listing: audit, rewrite, reviews and house rules.

Run your holiday-let listing: audit, rewrite, reviews and house rules.

Chase a late invoice (UK), with the interest worked out TESTED badge. Run on at least three real examples and scored against a written standard, with the record on file.

Chase a late invoice, from polite nudge to final reminder.

Chase a late invoice, from polite nudge to final reminder.

Check how AI search tools read your website TESTED badge. Run on at least three real examples and scored against a written standard, with the record on file.

Check if AI and Google can read your site.

Check if AI and Google can read your site.

Check a workflow file works across AI tools TESTED badge. Run on at least three real examples and scored against a written standard, with the record on file.

Check that a workflow file will load in other AI assistants.

Check that a workflow file will load in other AI assistants.
14 more: checked on real input before release, no badge yet

What we don't claim

  • Most of our cases are written by us or taken from public pages. They are cleaner than real work.
  • Most tests ran on Claude models. Three workflows were also run on ChatGPT's free model through the API, which can differ from the ChatGPT app.
  • Small tests. A difference of one or two outputs means nothing.
  • Store and host workflows were scored by hand by the team that built them, then checked by a second reviewer on our team.
  • No customer has rated anything yet.

Something wrong? More detail?

Write to the address on the About page with what you gave the workflow and what it returned. We reproduce it and change the workflow, and the fix goes to everyone who has it. The studies are under Research; changes to the free workflows are in the changelog.