Profiles by model

In short: Literal and structured, state the scope, use XML.

What is: Claude follows instructions literally: it doesn't generalize on its own or infer requests you didn't make. It responds very well to structure (XML) and explicit scope.

When to use: This guide's default. Great for structured extraction, coding and pipelines where you want predictable behavior.

When to avoid: Don't expect it to "guess" what you meant, be explicit.

💡 Tip: Avoid aggressive language ("CRITICAL: you MUST"), on recent models it causes tool overtriggering. Prefer "Use X when...".

In short: Objective + constraints + output contract; don't micromanage.

What is: Classic GPT performs with precise, outcome-oriented prompts. Reasoning models (o-series, GPT-5 reasoning) perform with a clear objective + strong constraints + output contract, WITHOUT prescribing every intermediate step.

When to use: Reasoning models shine on dense, long documents (contracts, financials) and complex decisions.

When to avoid: Do NOT instruct "think step by step" / manual CoT on reasoning models, it can hurt (they already reason internally).

💡 Tip: On GPT-5/o-series, raise reasoning_effort for hard tasks and give an explicit output contract instead of long examples.

In short: Persona · Task · Context · Format.

What is: Gemini performs best with the PTCF structure: Persona (who the model is and the audience), Task (the desired result), Context (background) and Format (how the output should look).

When to use: Any Gemini prompt, PTCF is the most reliable way to structure it.

When to avoid: It's not a straitjacket; treat it as a starting point and refine.

💡 Tip: Prompt engineering is iterative: start with PTCF and refine based on the responses you observe.