Claude Code's Leaked System Prompt Confirms the Rush Bias

Three weeks ago I argued that Claude Code has a rush bias: it is tuned to finish fast, not to finish well. Then the Claude Code source leaked, system prompt included, and the receipts are in the text. Instructions like "go straight to the point," "try the simplest approach first," "be concise," "keep text output brief and direct," "lead with the answer." Plus model annotations that patch specific model biases directly in the harness.

That is the finding. The rest of this piece is what it means for your work and what actually compensates for it.

What is the rush bias?

Rush bias, or action bias, is Claude Code's tendency to skip analysis, planning and verification and jump straight to execution. On simple tasks you barely notice. On complex tasks it produces sloppy work, because rushing through something complex always loses to taking your time. My read is it became more pronounced around 4.6 and later.

The complaints developers post match the pattern: glancing at images instead of reading them, treating "proceed, and do it correctly" as license to finish as fast as possible, inventing excuses, band-aid fixes over real solutions, and in nine out of ten PRs missing or skipping something that mattered. LLMs will also happily gaslight you about having done the work.

Why can't you just prompt it away?

Because the instruction fight is asymmetric. The system prompt gets a much larger share of the model's attention budget than anything you type. You can tell it to take extra time and be diligent, and on a long, complex task it loses that instruction as attention decays, while "be concise, lead with the answer" is stapled into every single turn. You basically cannot win from inside the conversation.

That is also the real lesson of the leak: the harness is the product, not the model. The same model behaves differently under a different system prompt. So the fix has to come from the harness side, and as far as I know Anthropic has not responded publicly. I might be wrong on that; I just have not seen it.

What you can do about it today

Three things help, none of them fully:

1. Use plan mode more. You should be doing this anyway; it forces the analysis phase the prompt tries to skip.
2. Use extended thinking on complex tasks.
3. Put diligence instructions in your CLAUDE.md so they are always present. Not guaranteed to work, for the attention-budget reason above, which is why you end up repeating them anyway.

The workflow that compensates: Claude plans, Codex executes

What I actually do now is run the two tools against each other's biases:

1. Claude drafts the plan, then I ask it where the weak spots are.
2. Codex reviews the plan, names what is lacking, and assigns priorities.
3. Claude, which holds all the context, works the sensible changes back in. A couple of rounds of this.
4. The finished plan goes to Codex for implementation. Codex does not have a rush bias; it just keeps going.
5. Claude checks Codex's work against the plan, and handles the small follow-up changes where diligence matters less.

You get Codex's literal-minded rigor and Claude Code's tolerance for vague instructions in one pipeline. The tradeoff is real on both sides: Codex is so literal that if you do not specify exactly what you want, you get something else entirely. The full division of labour is in the Claude Code vs Codex split.

Rush bias is real, it is in the system prompt, and the fix lives on the harness side. Acknowledge that first, then structure your workflow so something pedantic reviews everything something fast produces.