For ChatGPT 5.5, Claude Opus 4.7, Gemini 3.5 Flash, and similar current models
1. Start with the outcome, not a role
Avoid:
You are an expert strategist. Help me with pricing.
Use:
Create three pricing options. For each option, explain the customer fit, trade-off, assumptions, and recommended choice.
Newer models do not need as much persona setup. The transcript’s main point is that role-based prompting can backfire because newer models take wording more literally and may focus on the role instead of the task.
2. Say exactly what context to use
Avoid:
Write a proposal based on what you know.
Use:
Write a proposal using the discovery notes, brand voice guide, past proposal examples, and pricing sheet. If any source is missing, say so before drafting.
Do not assume the model will automatically inspect the right files, project knowledge, or pasted context. Name the materials it must use.
3. Define “good” clearly
Add success criteria:
Good output means: concise, specific, commercially useful, no invented facts, and formatted as a client-ready proposal.
For GPT-5.5-style models, OpenAI recommends describing what good looks like, which constraints matter, what evidence is available, and what the final answer should contain.
4. Use fewer process instructions
Avoid overloaded prompts like:
Think step by step, act as an expert, use first principles, critique yourself, be creative, be concise, be exhaustive…
Use:
Give the strongest answer. Explain key trade-offs. Flag uncertainty. Keep it under 600 words.
Latest models can plan more of the work themselves, so the prompt should steer the destination, not micromanage every internal step. OpenAI specifically warns that legacy prompts can add noise or make outputs overly mechanical with GPT-5.5.
5. Ask for an audit on complex tasks
For multi-file or multi-step work, add:
At the end, report what you processed, what you skipped, why anything was skipped, and whether the task is complete. Only say “complete” if every required item was handled.
This mirrors the transcript’s recommendation: as models handle longer tasks, users need a simple completion report to catch skipped files, transcripts, invoices, or subtasks.
Reusable prompt template
Task: [exact deliverable]
Use this context: [specific files, notes, links, examples, pasted text]
Goal: [what the output should help me do or decide]
Output format: [sections, length, tone, table/bullets/etc.]
Success criteria:
[criterion 1]
[criterion 2]
[criterion 3]
Constraints: [what to avoid, what not to assume, required citations, style rules]
Before finalizing, check whether every required source or item was used. End with a brief audit: processed, skipped, assumptions, and completion status.
Constraints: [what to avoid, what not to assume, required citations, style rules]
Before finalizing, check whether every required source or item was used. End with a brief audit: processed, skipped, assumptions, and completion status.Bottom line: for the latest LLMs, including ChatGPT 5.5, prompt less like “be an expert” and more like “here is the outcome, here is the context, here is what good looks like, and here is how to verify you finished.