CoreWise Academy

Foundations · Layer I / Practitioner

Claude and OpenAI, one prompting playbook

OpenAI and Anthropic prompting guides are one playbook when read side by side: lean out the prompt. Where they converge, where they diverge, which habits to drop.

Nº 018 · Vol. I·5 min read· Updated July 2026

Read firstBrief the model like a brilliant new hire (Assumes the prompting fundamentals so the lean-out advice reads as refinement, not first principles.)
“two vendors, one instruction: say less”

OpenAI’s latest-model guide for GPT-5.6 and Anthropic’s prompting best practices now open on the same instruction, phrased two ways: lean out the prompt. The habits you built to wrestle older, weaker models are the exact habits both vendors are now telling you to unlearn.

01Where they converge

OpenAI’s core advice is blunt:

Say explicitly which actions the model may take on its own and which need a human, so it stops pausing where it shouldn’t while still blocking anything external, destructive, or costly.

Claude’s guide lands in the same place from the other side. It tells you to soften the aggressive forceful language meant to stop the model doing too little, the “CRITICAL, you MUST use this tool” phrasing that current models over-respond to, down to plain “use this tool when.” And it warns that these models overengineer by default, reaching for extra files and unrequested abstractions, so a minimalism instruction earns its place. Calibrating that language up and down is the whole subject of Stop shouting at the model.

Both diagnoses share a root cause. The emphasis and repetition you learned were compensating for models that under-followed instructions. Current models follow them literally and weight an over-emphasized rule too heavily, so the compensation now backfires.

“Lean prompts correct for models that got better at following instructions while your habits stayed the same.”

02Where they genuinely diverge

Convergence on the prompt doesn’t mean agreement on the underlying systems. Three splits are real:

SplitOpenAIClaude
Reasoning controlNumeric ladder: reasoning.effort none → max, plus mode “pro”Adaptive thinking calibrated by an effort parameter
Coordinating toolsProgrammatic Tool Calling runs server-sideOrchestration stays in the client loop or code execution
Reasoning reuseEncrypted reasoning via previous_response_idThinking blocks re-sent in conversation history

Reasoning control. OpenAI exposes a numbered scale, reasoning.effort from none through max, plus a reasoning.mode “pro” switch that spends more model work on one request before returning a single answer. Claude runs adaptive thinking calibrated by an effort parameter and query complexity, with the model deciding depth per turn rather than you picking a step. That effort parameter also controls how hard the model thinks, which Guardrails for hours-long agent runs puts to work.

Coordinating tools. OpenAI’s Programmatic Tool Calling lets the model write code that calls tools server-side, process several results, and hand back one small structured result. It’s meant for bounded filter, aggregate, and validate work, not decision-dependent sequences. Anthropic’s similar path keeps the coordinating in the app’s own request loop or its code-execution tool (our own inference, not a claim on either page).

Reasoning reuse. OpenAI carries encrypted reasoning across turns with previous_response_id plus reasoning.context, so prior thinking can persist when goals stay stable. Claude re-sends thinking blocks in the conversation history instead (also our own inference). Same goal, opposite internal wiring, and the integration code is not portable between them.

03How much to trust the numbers

OpenAI reports internal tests where leaner prompts scored roughly 10-15% higher while cutting tokens 41-66% and cost 33-67%. Those numbers are the strongest argument in the guide, and they are also the softest. They’re self-reported, with no methodology, task mix, or baseline stated, so you cannot tell whether they hold on your workload or on a benchmark shaped to flatter the advice.

What makes the direction trustworthy is that the other vendor moved the same way independently. Two companies telling you to cut repetition, from different testing setups and different model architectures, is stronger evidence than either vendor’s percentage.

“Take the direction as settled and the numbers as an invitation.”

Run the before-and-after on your own prompts and measure the difference yourself, because that is the only number that describes your system.

Further reading

FAQ

Both vendors now tell you to cut something from prompts you were previously told to add. What is it, and why the reversal?

Repetition and heavy approval-gating. Older guidance leaned on "CRITICAL, you MUST" emphasis and scattered "ask first" phrases to force behavior on weaker instruction-followers. Current models over-react to that language, pausing when they should act, treating an emphasized rule as louder than the others. OpenAI says state each instruction exactly once and define autonomy in one place; Claude says soften the aggressive forceful language meant to stop the model doing too little. Same fix, two companies.

Where do the two vendors genuinely diverge, beyond a difference in wording?

Coordinating tools. OpenAI Programmatic Tool Calling has the model write server-side code that calls tools, processes several results, and returns one small structured result, good for bounded filter/aggregate/validate work. Claude keeps the coordinating in the app's own request loop or its code-execution tool. That is a real architectural split. Reasoning reuse diverges too: OpenAI carries encrypted reasoning across turns via previous_response_id plus reasoning.context, where Claude re-sends thinking blocks in the conversation history.

OpenAI reports leaner prompts scoring 10-15% higher while cutting cost 33-67%. How much weight should that carry?

Read it as a rough direction worth checking. The numbers are self-reported internal tests with no methodology, task mix, or baseline disclosed, so you cannot check whether they transfer to your workload. Treat them as a reason to run the experiment yourself, not as a measured result to quote. The direction is backed up by the other vendor moving the same way; the magnitude is not something you should bank on.

Sources

Original work by the editor, no video source.