Blog ·
I Removed Every Codex Skill. The First Homepage Was Better
I removed every Codex skill and gave GPT-5.6 a clear task contract plus a verifier. The first homepage was better, but one strong result did not settle the argument.
For months, I had been teaching my coding agent how to design. Every mistake became another rule. The design skill kept growing.
Then I watched Boris Cherny talk about removing more than 80% of Claude Code's system prompt. I wanted to see what GPT-5.6 would do without my accumulated guidance.
So I tried it. The first homepage mockup was far better than I expected. My reaction was embarrassingly simple: wow.
One run proves little. It still made me question whether all those instructions were helping.
What Boris actually said
The 80% figure referred to one Claude Code system prompt during the move to Opus 5. It wasn't a claim that 80% of all agent instructions are useless.
His advice was to remove old scaffolding, use the new model on real work and watch what goes wrong. When a failure repeats, add back the smallest prompt, skill or tool that fixes it.
Boris still gives the model a task, guardrails and exit criteria. He calls verification one of the most frequently missed parts. The long recipe telling the model exactly how to work can go.
You define the destination, the boundaries and the evidence. The agent chooses the route.
A better structure for prompts
Shorter cannot mean vague. The prompt still needs to give the agent a clear contract:
| Section | What belongs there | Dori example |
|---|---|---|
| Task | The job the agent must perform. | Create a new homepage mockup. |
| Context and inputs | Facts the agent cannot know: the customer, offer, true claims, source files and available assets. | Use the existing Dori business context and client work. |
| Expected outcome | What should exist afterward and what it should achieve. | An English homepage that explains the offer to a small-business owner and leads to booking a call. |
| Scope and guardrails | Hard boundaries, protected files, forbidden actions and decisions that require approval. | Do not invent clients or testimonials. Do not publish the page. |
| Done when | Observable conditions that allow the agent to stop. | Required content is present, calls to action work and the page is responsive and keyboard accessible. |
| Verification | The checks that prove those conditions were met. | Render named viewport sizes, click every link, test keyboard navigation and inspect the screenshots. |
The outcome sets the direction. Done when turns it into facts you can observe. Verification explains how the agent will collect proof.
"Do not invent client testimonials" is a boundary. "Use three columns with 24-pixel gaps" is usually an implementation choice. Unless it is a real requirement, leave it to the agent.
Where Karpathy fits
Andrej Karpathy's description of context engineering widens the idea beyond the prompt. The agent needs the right files, tools, state and feedback, without unrelated material crowding the context window.
In his discussion of agentic engineering, the human owns the specification, design and quality bar while the agent handles implementation. His autoresearch agent has broad freedom inside a narrow experiment with a fixed metric. That verifier makes the autonomy practical.
My first test
I tested the idea on a real homepage project without my accumulated design guidance.
The result was surprisingly good. Much better than I expected from a cold run.
I am not deleting every skill forever. My new default is a clear task contract plus a verifier. A permanent instruction comes back only after the same failure comes back too.