New AI model? Delete everything first, then set up again

A year ago, AI development followed a simple rule: the more context, the better. A long AGENTS.md, a CLAUDE.md, a DESIGN.md, dozens of skills, hooks for every case and guardrails in capital letters. Whoever explained a lot got better results.

With the 2026 models, that flips. The best setup for a new model often starts with deleting. Then you test what the model can do on its own and rebuild only what is really missing.

What an AI setup used to look like

A typical project with AI agents had these building blocks in 2025:

• Context files. AGENTS.md or CLAUDE.md with a project overview, folder structure, commands and style rules. Often several hundred lines.
• Design and text rules. Separate Markdown files for colours, spacing, tone and forbidden words.
• Skills. Ready-made instructions for recurring tasks, from commits to deployment.
• Hooks. Scripts that run before or after every action of the agent and step in.
• Guardrails in the prompt. “CRITICAL: You MUST always …”, “Think step by step”, “Check everything three times”.

There was a reason for it. The models forgot conventions, skipped tests or built things nobody had asked for. Every rule was an answer to a real mistake.

Why this slows down new models

New models often no longer make the mistakes these rules were written against. The rules stay, though. And they keep working, only now in the wrong direction.

New models follow more closely

Older models overlooked instructions. So people wrote them loudly: in capital letters, with “ALWAYS” and “NEVER”. New models listen closely. In its prompting recommendations for the Claude 4 models, Anthropic already writes that such wording now leads to overeagerness. A tool that “CRITICALLY” must always be used then gets used even where it doesn't fit. The recommendation: phrase it calmly, for example “Use this tool when …”.

Old crutches contradict new abilities

Claude Opus 5.5 thinks before every answer and chooses the depth itself. A line like “Think step by step” is therefore redundant. According to Anthropic, answers start faster without this line, with no noticeable loss of quality. The same goes for “Be brief” or “Check everything three times”: Anthropic recommends retesting such corrections on the new model instead of simply carrying them over.

More context costs more and achieves little

In February 2026, researchers at ETH Zurich examined whether AGENTS.md files really help coding agents. Among others, Claude Code and Codex were tested. The result: context files barely improve the success rate. Files generated by AI even make it slightly worse. Costs, on the other hand, rise by more than 20 percent on average because the agent takes more steps.

Particularly revealing: the agents follow the instructions in the files well. So they do what the file says. It is just that much of it doesn't help. An agent that can read the repository itself doesn't need an overview of the folder structure. According to the study, the files are mainly useful for unusual conventions the model cannot guess.

What the vendors do themselves

Boris Cherny built Claude Code at Anthropic. In an interview with Y Combinator in July 2026, he said his team removed around 80 percent of Claude Code's system prompt without the results getting worse. His advice to everyone working with AI agents: every six months, delete your CLAUDE.md, your skills and your hooks. Then see what the model does without them.

That is remarkable. The people who build the model treat their own setup as a consumable. A harness, meaning the shell of prompts, tools and rules around a model, only lasts until the next model.

How we set up again after a model change

1. Secure what exists. The old setup goes into its own branch. We delete in the working copy, not in the archive.
2. Switch everything off. Context files, skills, hooks, guardrails in the prompt. What remains is the bare model with access to the code.
3. Test with real tasks. Five to ten typical everyday tasks: a bug fix, a new page, a restructuring, a text change. We watch and note every mistake.
4. Bring back only what keeps going wrong. A single slip isn't worth a rule. A mistake that happens twice is. The rule is worded as narrowly as possible, calmly and with a reason.
5. Put anything checkable into code, not text. What a linter, a test or a type check can catch belongs there. A failing test convinces any model. A rule in a Markdown file doesn't always.
6. Note the date. With the next model, the round starts again.

Which rule stays, which goes

What may stay

Deleting doesn't mean working without context. What stays is what the model cannot read from the code:

• Business rules and customer wishes that appear nowhere in the code.
• Unusual conventions that go against the standard.
• Commands for testing and checking, if they are not obvious.
• Limits for anything involving money, customer data or production systems.

The last point needs more than text. Security limits belong in permissions, sandboxes and approvals. A rule in a Markdown file doesn't stop an attacker who slips new instructions to the agent through an email or a web page. More on this in our article on agentic attacks. And fewer hooks and skills also means fewer files in which malicious code can hide. How attackers use exactly such configuration files is covered in the article on agentjacking.

Frequently asked questions