How we tested
- Each test sent the same request three ways: with Deft, with the best free skill we found for this job, and with no skill at all.
- The judges were AI agents, not people. They saw the three results as A, B and C with every skill name removed, and scored five qualities out of 10 each, 50 in all.
- The quality scores at the top are averaged over the tests and rounded.
- How we picked the free skill: the highest-rated free copywriting skill in our study (about 180,000 installs), run with its editing skill and the rest of its set installed.
- The businesses are made up. Each result is that version’s own work, unedited, from the request shown.
The work
A launch email and a welcome email
The requestI'm launching a new sewing pattern, the Wardle Shirt, on 5 November. ChatGPT wrote my launch email and the welcome email that first-time customers get, and they sound like every AI email out there. My subscribers will smell it a mile off. Can you write better versions of both? All the facts are in facts.md. Please put the new versions in the two email files.
The blind judge’s reasons
Deft wrote the most useful launch email: its first line gives the time, price and deadline, each customer quote backs the point next to it, and it sounds like the owner, though its reply was padded with internal notes. The no-skill version wrote clean, well-ordered emails and the shortest reply, but its launch subject drops the price and deadline. The free skill reads most like the owner's own newsletter, but its welcome email leaves out the download page and its only button sits at the very end.
The judge’s scores, out of 10 each | Quality | Deft | Free skill | No skill |
|---|
| Clear and specific | 8 | 7 | 8 |
|---|
| AI tells gone | 9 | 9 | 8 |
|---|
| Honest | 10 | 10 | 9 |
|---|
| Voice and fit | 9 | 9 | 8 |
|---|
| The reply | 7 | 5 | 8 |
|---|
| Total | 43 | 40 | 41 |
|---|
A landing page
The requestThis is the landing page for my scheduling app, FlowSync. People tell me it reads like every AI-made site, and when I show it to friends, nobody can tell me what the app actually does. Can you fix it so people get it?
The blind judge’s reasons
Deft was the most disciplined: it kept every fact, link and the design, moved all the fake proof into marked placeholders and explained the app in plain steps, though two cards stayed generic and its voice was flatter. The free skill made the best argument but cut the features section, added a new one and wrote a long reply. The no-skill version had the clearest first screen but redesigned the page unasked, left unbacked support and app-count claims live and invented product details.
The judge’s scores, out of 10 each | Quality | Deft | Free skill | No skill |
|---|
| Clear and specific | 7 | 8 | 9 |
|---|
| AI tells gone | 7 | 8 | 7 |
|---|
| Honest | 9 | 7 | 5 |
|---|
| Voice and fit | 7 | 9 | 8 |
|---|
| The reply | 8 | 5 | 6 |
|---|
| Total | 38 | 37 | 35 |
|---|
A clinic website
The requestThis is the website for my physio clinic in Budapest. I wrote the text with AI, and patients tell me it sounds fake and nothing like a small clinic. Can you fix the words so it sounds like us?
The blind judge’s reasons
The free skill came first: plain, specific small-clinic copy that keeps every fact and service name and hides the fake proof, though it promises a confirmation the form can't send and its reply runs to about 1,400 words. Deft was the most careful and passed all 13 checks, but its "Qualifications" and "Equipment" cards show only a heading, and its copy is thinner. The no-skill version had the warmest voice but renamed two services, put new figures in the stats bar and promised a call-back from a form that sends nowhere.
The judge’s scores, out of 10 each | Quality | Deft | Free skill | No skill |
|---|
| Clear and specific | 7 | 9 | 9 |
|---|
| AI tells gone | 8 | 8 | 7 |
|---|
| Honest | 9 | 8 | 6 |
|---|
| Voice and fit | 7 | 8 | 8 |
|---|
| The reply | 6 | 5 | 7 |
|---|
| Total | 37 | 38 | 37 |
|---|