One name on every guide, every page dated, every number sourced.
Your agent will run the commands these pages print. This page shows who stands behind them, how each one is checked, and what happens when one is wrong.
You already ship with agents.
The question is which level you ship at. The free scorecard places you on the ladder, and the free ladder pages show the next rung.
$59.99 once for 12 months of access. Nothing renews on its own. 30-day refund on a first purchase.
One developer at Level 4 isn’t a team at Level 4.
The hard part isn’t adopting agents. It’s moving the whole team up a rung together: one harness, one loop, one review standard. The Team Toolkit covers up to 10 people for 12 months, on one invoice: you plus up to 9 people you add by email address.
$299.99 once for up to 10 people and 12 months, on one company invoice. 30-day refund on a first purchase.
You’re betting the roadmap on agents. Know which level you’re betting from.
Your engineers already run agents. What nobody can answer yet is at what level, behind which gates, and what it did to review time and incidents. Fixed-price services for teams, delivered personally by Michał Jaskólski, start with a short call.
No obligation. We’ll look at how your team builds today.
Engineers at these companies work with AI Developer Toolkit
- Grupa Morizon-Gratka (opens in a new tab)
- Ringier Axel Springer (opens in a new tab)
- TasteRay (opens in a new tab)
TasteRay is also where the author is CTO.
Agents write the code. You still sign for it.
Andrej Karpathy describes what stays with you: “you are in charge of taste, engineering, design, and whether the system makes sense.” AI Developer Toolkit teaches both halves of that job: handing over bigger tasks, and checking what comes back.
“You are still responsible for your software, just as before.”
How much code agents write, and who checks it
Three measurements, each with its publisher and date
51%
of code is AI-authored on average, across more than 400 companies
48%
of developers always verify AI code. 96% do not fully trust it.
Michał Jaskólski answers for every guide on this site.
He builds with agents every day as CTO at TasteRay and teaches AI as a lecturer at SWPS University. The guides record the patterns that survived a review. They also say where an agent is the wrong tool, and where a human still has to look.
Why a named author matters to you. When a guide is wrong there is one name to hold to it, and the page’s date shows when it last changed.
-
2000
Shipping web apps
26 years of building for the web. Two IPOs along the way.
-
2023
Cursor, from day one
Prototypes came easy. Production did not: reviews, edge cases, scale. The official docs covered features and YouTube covered demos.
-
2025
CTO at TasteRay
Building the whole product with agents, and writing down every pattern that survived a review.
-
Now
AI Developer Toolkit
Those notes became 400+ guides and 100+ recipes, each showing the date it was last updated.
How a guide is checked.
The guides are made with the method they teach: an agent drafts from primary sources, a check against the installed tool follows, and a rubric with a bar of 9 decides what ships. The drawing shows the order and which steps you can see for yourself.
One guide, from the installed tool to your screen.
Followed through with one real example: a model suggests a Codex flag that does not exist.
- Primary source
- Done before publishing
- Visible to you
-
Primary source
--help
The installed binary’s own
--help, with the version it reports written down. The vendor’s live docs are the second source.codex-cli 0.160.1
Primary source
-
Draft
From the sources
An agent drafts the guide from the two primary sources, under the style guide.
A model suggests
--approval-modeBefore publishing
-
Check
Against the tool
Each command, flag and setting is compared with the binary’s
--help, then the live docs. A claim that something does not exist names the version it was checked on.Not in
--help. The real flag is--ask-for-approvalBefore publishing
-
Review
9 / 10
The page is scored on ten criteria. The bar is 9, and below it the page goes back for another pass.
Before publishing
-
Publish
Last updated
The page goes live showing its date. The Setup Pack’s catalogue also names the release each entry was checked on.
Checked 2026-10-10 on 0.160.1
You can see it
-
Report
Improve this page
On a wide screen, every guide has an Improve this page button. It opens a GitHub issue with the page’s address already in the title.
A reader reports a flag that fails
You can act on it
# what a model suggested codex --approval-mode never # what Codex CLI 0.160.1 accepts codex --ask-for-approval never
- Drawing
- The order of checks on one guide
- Applies to
- Claude Code, Codex, Cursor
- Note
- An illustration of the process with one real example, not a measurement
Every number names who said it, and when.
Five rules decide what gets printed. The specimen shows what travels with a number.
Anatomy of a number
+441%
Median time a change waits in review
-
Third parties only.
The site publishes no outcome numbers of its own. Every statistic comes from a publisher you can look up, with the date it was published.
-
Output and quality, together.
Faros AI reports +66% epics per developer and +242% incidents per pull request. The site never quotes the first without the second.
-
Evidence against the thesis is printed too.
METR (2025-07-10) found that experienced developers’ tasks took 19% longer with AI. METR calls its 2026 follow-up an unreliable signal, and the site says that as well.
-
Every share names its owner.
A vendor’s figure about its own code is labelled as that vendor’s, never as the industry’s. The site uses no speed multiplier, because none of the studies it cites measures one.
-
Credentials say who checks.
The author’s 26 years are never offered as proof of your result. There are no testimonials here.
Written for four readers.
The same method at four scales. Each role has a reading path and a free scorecard, so you can check the fit before you pay.
For engineering teams
Want the whole org on one level?
Fixed-price services, delivered personally by the author. They are on the Services for teams page.
Free · 25 questions · ~8 minutes
Which level is your AI workflow at?
Your place on the ladder (Levels 1–4) and the one move that takes you to the next rung.
The company behind it.
AI Developer Toolkit is published by Wondel.ai sp. z o.o., a company registered in Warsaw, Poland.
- Publisher
- Wondel.ai sp. z o.o.
- Address
- Twarda 18
00-105 Warsaw, Poland - Registration
- KRS 0001029516
NIP 5252951298
REGON 524989057 - Payments
- Polar is the merchant of record and sends a VAT invoice for every charge.
Publisher profiles
- X (@devtoolkit_AI) (opens in a new tab)
- GitHub (wondelai) (opens in a new tab)
- Discord (opens in a new tab)
- Wondel.ai (opens in a new tab)
Questions about any of this? Choose the right address on the Contact page.
Check a page. Then buy the Toolkit.
Read a free guide in the Ladder section and take the AI Dev Score. If you then buy the Toolkit and it does not change how you work, ask for a refund within 30 days.
$59.99 once for 12 months of access. Nothing renews on its own. 30-day refund on a first purchase.
What to look for on that page
- The last-updated date near the top.
- A version number beside any claim that a flag or command does not exist.
- A publisher and a date beside every statistic.