Are Codex and Claude Code really coding tools?

I call these tools knowledge super tools. Yes, they write code and build websites. More valuable is that they take over more and more of the work you would otherwise do yourself at the computer: emails, spreadsheets, documents, recurring tasks. To start, you need a process you can describe cleanly. Coding knowledge appears nowhere on the list. codex and claude-code are the best known examples. Searching an inbox and drafting a reply. Filling a spreadsheet by clear criteria. Testing a website and reproducing a bug. Turning a transcript into a recap. Every one of these tasks can be broken into checkable steps, with a human approving at the end. For that to hold, the tools need a place where your knowledge lives: a second brain of notes, decisions and documents that both you and the tool can reach. I took the method behind this from the entrepreneur Alex Hormozi. It comes down to getting very precise about what you actually want done. Picture hiring a video editor. The real steps then read: read the brief, topic and research, run an interview, check the facts, plan the structure, write the draft, build the graphics, sharpen the description for search, gather feedback, revise, get approval, publish, review the performance, archive it cleanly. From there the question becomes: which sixteen activities does this role really do?

How to break one role into mini workflows

I took the method behind this from the entrepreneur Alex Hormozi. It comes down to getting very precise about what you actually want done. Picture hiring a video editor. The real steps then read: read the brief, topic and research, run an interview, check the facts, plan the structure, write the draft, build the graphics, sharpen the description for search, gather feedback, revise, get approval, publish, review the performance, archive it cleanly. From there the question becomes: which sixteen activities does this role really do?

Only after that do you decide where a person sits, where an agent works, and where a plain automation is enough. The difference is quickly told: an agent makes free decisions, an automation follows a fixed chain. I went through this breakdown for my own companies once, linearly, task by task. That is the tedious part, and it is the part that carries everything later.

  • The trigger: when it should happen.
  • The input: what you hand over.
  • The context: what it has to know about you and the matter.
  • The action: what exactly gets done.
  • The destination: where the result goes and in what form.
  • The quality criteria: how a good result is recognised.
  • The check: who or what verifies it before the result counts.

Five things I do with them that are not code

  • Pull everything on an insurance matter out of the inbox and draft a reply, started from my phone while the computer sits at home.
  • Prepare invoices from a spreadsheet. My assistant checks and sends them. Her time goes into checking, and that saves real money.
  • Run three research jobs every Tuesday at noon. The result only enters the spreadsheet if it meets every quality standard.
  • Have the form for the Salzburg tourism levy pre-filled. I submitted it myself. I used to spend hours on that task.
  • Turn a transcript into a recap, checked against the criteria of the series.

The quality gate

A quality gate is a list of questions that have to be answered before a result is passed on at all. In my setup there are ten. If a draft fails them, it goes back into the loop until it passes. Since these gates exist, the whole setup runs a lot calmer. They are also the reason I can look at results properly: whatever fails the gate never reaches my desk.

Where the human keeps the approval

Standard permissions are enough to start. The tool works inside the folder you opened, and it asks as soon as it wants to go beyond it or delete something. Every project folder is a small sandbox, an invisible wall that only opens when you explicitly agree. The second level, automatic review, confirms a lot on its own and asks less often. Full access is something I recommend to very few people, and I only use it because this machine is set up for exactly that. The same rule sits above all three levels: publishing, sending, paying and deleting need the approval of a person.

Even when I say send this email to this person, it does not do it right away.

The agent then answers along these lines: you told me to send, my rules say I check with you first anyway. Here is the email I intend to send. Only after my yes does it go out. Why access to your own data beats a better worded instruction is in Why access beats the prompt. Which model belongs behind which step is sorted in Who thinks, who works, who just runs.

  • March 2026: The switch to Codex. Before that, most of it ran through Claude Code.
  • 1 June 2026: The article about the method behind this series is itself drafted with Claude Code and edited afterwards.
  • June 2026: The Deep Dive shows the setup live, with 17,729 local threads since 15 March in the profile.

On record

KI DeepDive Codex Agent Setup, public live session, June 2026.

  • Workflow patterns from the session: email research, spreadsheet work, browser checks, documents, forms, a dedicated agent setup
  • Every workflow is first broken into role, process, mini workflows and a quality gate
  • A quality gate is ten questions a result has to answer before it moves on
  • Three permission levels: standard, automatic review, full access. Every project folder is a sandbox
  • Publishing, sending, paying and deleting only happen after a human approves
  • 17,729 local threads since 15 March 2026, among them one task that ran for 8 hours 39 minutes

This note keeps growing

2026-09-03: Deepened: the breakdown method with the video editor example, the seven entries per activity, five tasks without code, the ten questions of the quality gate, the three permission levels, timeline and source list.

2026-09-02: Planted from the Codex Agent Setup Deep Dive.

ZukunftBilden GmbH · Salzburg · +43 681 81655313 · office@zukunftbilden.eu