Work with me
English, Choose language
Menu
English, Choose language

If your AI agent, the software that does work for you on the computer, also said “Fixed, everything works,” and the problem is still there, you're not alone.

Could the problem be something other than the agent? It didn't lie. Nobody asked it to prove anything. “Done” is the natural end of every task, so it shows up even when nothing was run.

It wrote the change, and to the agent writing feels like doing. Or it trusts an old result, or checks a stand-in instead of the real thing. Every time, the outcome is the same: you find the problem, not the agent.

And every fake “done” like that, found two days later, takes you hours.

Axel Jedlitschka

It all started with an income and expenses table I wanted to build for myself in Excel. After a few versions I moved to Google Sheets, and from there it kept growing.

While booking sales meetings by phone, I started building systems and automations, processes that run on their own, for business owners. I built a system for business forecast analysis, and it served the company's marketing and sales very well. A system for a business owner is worth what it actually does for them, because writing is not running.

Axel Jedlitschka

  • Evidence, not a statement

    Every “done” comes with its evidence: what was run, and what came out.

  • Fresh proof

    Know the proof came from a run after the last change, not before it.

  • Numbers that add up

    Every percentage with its denominator in the same row, and every number with its source.

  • An attempt to break it

    Know the agent tried to find how it is still broken, and what it deliberately did not check.

  • One line on top

    Know from the first line whether it is done and verified, blocked, or not finished yet.

The rule is simple: no task is done without proof. Proof is something you could check yourself, produced by a real run after the last change. Before the agent writes “done”, it runs four checks.

  1. Freshness.The proof comes from a run after the last change. Whatever printed before the edit, in an earlier conversation or from memory, is stale.
  2. Concrete evidence.Every verdict, pass or fail, comes with its evidence: the command and what it printed, or the line from the file with its location. “Verified” with nothing under it is a statement.
  3. Numeric consistency.Every number with a source, and every percentage with its denominator in the same row. If it says four files changed, the list under it has four lines.
  4. Contrarian pass.One pass whose only goal is to find how it is still broken: the edge case, the second path, and whatever sits right next to the change. Nothing turned up? It says what it tried, and what it deliberately did not check.

“I wrote the code” is not proof. Writing is not running.

Verification kit · a kit by Axel Jedlitschka

Before the agent says it is done, it checks that it really happened.

ZIP · Skill + hook

  • The installer checks its own work: it compares every copied file with the original and prints how many match
  • It shows what it will do, and asks before writing anything
  • The rules work with any agent that reads an instruction file
  • A skill for Claude Code: an instruction file the agent loads when it is about to report a task as finished
  • The four checks: freshness, concrete evidence, numeric consistency and a contrarian pass
  • A one-line status on top of every reply: done and verified, blocked, or not finished
  • Four before/after examples: a code change, a data report, a content task and a deployment
  • A rules block ready to paste, for any other agent that reads an instruction file, such as Codex or Cursor
  • Install in one command, with a backup of any earlier version
  • An optional reminder (a hook): one line each time the agent finishes a turn

What's in the ZIP

One ZIP file · one command to install

  • A skill for your agentInstructions your agent loads when the task comes up
  • 1 hookSmall actions that run on their own at set moments
  • One-command installAsks before it writes, and backs up any older version
  • Also for Codex, Cursor and Gemini CLIA ready-to-paste instruction file for any agent
  • PDF guide, 7 pagesIn the language you pick on the form

No cost

Verification kit

What you get
The command and what it printed, or the line from the file
When it was checked
In a run after the last change
Who finds the problem
The agent looks for it, in the contrarian pass

“Done” without proof

What you get
“Fixed, everything works”
When it was checked
In a test that passed before the change
Who finds the problem
You, not the agent

A skill, four checks, a status line, a rules block for any agent, and a guide of 8 pages

Price: No costNo credit card, and no part is held back for payment

Get the kit

The kit is yours right away, at no cost. You only get marketing messages if you tick the box, and you can unsubscribe at any time.

As soon as you leave your details, a thank-you page opens with a button to download the kit. You choose the guide's language in the form: Hebrew, English or Spanish.

By Axel Jedlitschka

Promised

  • The full kit, at no cost, as soon as you leave your details.
  • No part is held back for payment.
  • Your details are not sold, and are not passed to anyone except the infrastructure providers that run the site.
  • A question about the kit? Send a message on Instagram to @theonlyaxelj.

Not promised

  • A result. The kit teaches the agent to prove its “done”, and the result depends on how you apply it and on your business. It is general knowledge, not personal advice.
  • That you stop checking. Your check gets short, but it is still yours.

There is no purchase and no payment here, so there is nothing to cancel. You can withdraw your consent to marketing messages at any time.

How do I get the kit?

You leave your details, and a thank-you page opens right away with a button to download the kit.

Is there really no cost?

Yes. No credit card, and no part is held back for payment.

I'm not a programmer. Can I install it?

Unzip the file you downloaded, and from the folder it creates, run one command in the terminal, the window where you type commands to your computer: bash install.sh. The script shows what it will do and asks before writing anything. The default answer is “no”.

Doesn't this slow the agent down?

A little. Re-running a check takes seconds. Discovering a fake “done” two days later takes hours.

Can't the agent just make up the proof?

An agent can write anything. That is why the rule asks for specific evidence: a command you can run, a file and line you can open. When the evidence is specific, checking it takes you ten seconds.

Does this only work with Claude Code?

The skill and the reminder are built for Claude Code. The rules themselves are in the AGENTS.md file, and they work with any agent that reads an instruction file, such as Codex or Cursor.

Will I get spam or phone calls?

Your details are used to send you the kit and to follow up on it. Marketing messages are sent only if you ticked the consent box, and you can unsubscribe at any time. You can also ask to see, correct or delete your details.

Where can I get help?

A question about the kit: send a message on Instagram to @theonlyaxelj.

You can keep opening the app after “Fixed, everything works,” or you can get the “done” with the proof already inside the reply.

A kit to install and a PDF guide of 8 pages: four checks, a status line and a rules block for any agent, at no cost.

Price: No costNo credit card

Get the kit

The kit is yours right away, at no cost. You only get marketing messages if you tick the box, and you can unsubscribe at any time.

The kit is ready to download as soon as you leave your details

The kit is yours right away, at no cost. You only get marketing messages if you tick the box, and you can unsubscribe at any time.

Your details are used to get back to you. What is collected, why, and who receives it: see the Privacy Policy (opens in a new tab) By sending, you agree to the terms (opens in a new tab)