Skip to content

Pick your style: Denkbaar

Book a conversation

Articles

Knowledge

An agent that does what you want, every time

Consistent work from an agent does not come from a better prompt, but from fixed rules, a check and a learning loop.

Two people compare a laptop result with a card and a checking panel.
Fixed rules, checks and learning from mistakes make agent work more consistent. Illustration: Denkbaar.

An agent that does good work one day and not the next is not usable. You build consistency in, with rules, checks and a way to learn from mistakes.

Plain technical English An agent follows the rules in its context. A rule that must hold becomes an automatic check. Each mistake becomes a new rule or a new check.

Why it matters

A language model does not give the same answer every time. For a text, that is sometimes fine. For a quote, an invoice or a deploy, it is not. If you must check every step yourself, the agent saves you no time.

Three documents lie beside a sample card and one has a different orange block.
For a quote or invoice you want the right result every time. Illustration: Denkbaar.

How it works

noyes

Task with done criterion

Agent works

Short rules and skills

Check green?

Fresh reviewer

Mistake? New rule or check

The agent works until the check is green; a fresh reviewer looks; every mistake becomes a rule.

The Claude Code guide gives the same building blocks: a short instruction file, procedures in skills, a check that Claude runs itself, and hooks for rules without exceptions. A hook is a piece of code that runs automatically, for example before each commit.

A reviewer places an orange card with a rules folder beside a screen.
Turn a repeated mistake into a fixed rule or automatic check. Illustration: Denkbaar.

What you do

  1. Say when it is done. A test that must pass, or a screen that must look right.
  2. Write few rules, and keep them short. Long instructions are followed less well.
  3. Put ways of working in skills. We use, among others, test-driven-development, systematic-debugging and verification-before-completion from obra/superpowers (MIT).
  4. Make hard rules automatic. At Denkbaar, a check blocks every commit that is too large.
  5. Let someone else check. A fresh agent that did not see the build sees more.
  6. Learn from every mistake. One line in a lessons file. If it happens a second time, it becomes a fixed rule or check.

What we can do with you: we set up an agent that does the same good work for your task every time, with the checks included. Book a conversation.

Sources

  1. Claude Code Docs: Best practices for Claude Code
  2. obra/superpowers (skills for coding agents, MIT)
  3. Anthropic: Building effective agents

Topics: agentsmethodquality