Prompt Patterns

A pattern is a reusable skeleton you fill in. The fifteen below cover analysis, creation, editing, research, and multi-step work across Claude, ChatGPT, and Gemini. Copy a template, swap the bracketed slots, and paste it into the chat.

Pattern Index

Foundation Patterns

1. Role + Task + Format

When to use: your default opening move for any non-trivial request.

You are a [role with relevant expertise].
Your task is to [specific task].
Respond as [a table / numbered list / 3 short paragraphs].
Constraints: [tone, length, what to avoid].

2. Context Dump

When to use: the model keeps guessing for lack of background. Front-load the facts, then ask.

Background you need:
- [fact 1]
- [fact 2]
- [constraint or goal]

Now, given all of that: [your actual request].

Save recurring background once in a Claude Projects custom instruction or ChatGPT Custom Instructions, so every chat starts with it.

Analysis Patterns

3. Structured Analysis

When to use: you have a long document or dataset and want consistent findings, not a vague summary.

Analyse the [document/data] I provided.
Use exactly these headings:
1. Summary (3 sentences)
2. Key findings (bullets)
3. Risks or gaps
4. Recommended next step
Quote the source for any specific claim.

4. Compare-and-Score

When to use: choosing between options and you want a defensible ranking.

Compare these options: [A], [B], [C].
Score each from 1-5 on: [criterion 1], [criterion 2], [criterion 3].
Present as a table, add a one-line verdict, and flag the closest call.

5. Extract-to-Table

When to use: turning unstructured text (emails, notes, a PDF) into rows you can sort or paste.

From the text below, extract every [item type].
Columns: [field 1] | [field 2] | [field 3].
One row per item. Write "unknown" where a field is missing.
Output only the table.

For real spreadsheets, upload the file and let ChatGPT Code Interpreter parse it.

Creation Patterns

6. Draft-and-Outline

When to use: longer writing where you want to approve the skeleton before the prose.

I'm writing [piece] for [audience], about [length].
First, give me an outline only.
Wait for my approval, then draft section by section.

Ask Claude for the draft as an Artifact to iterate in the side panel, or use ChatGPT Canvas to edit inline.

7. Few-Shot Template

When to use: you need a precise format and describing it isn't working — so show it.

Here is an example of the exact format I want:
[paste one finished example]

Now produce the same thing for: [new input].
Match the structure, length, and tone exactly.

8. Persona Voice

When to use: the tone and audience are the hard part, not the facts.

Write [piece] in the voice of [persona/brand].
Audience: [who]. Reading level: [level].
It should feel [adjectives]. Avoid [words/clichés to skip].

Editing Patterns

Editing is sharpest in a live surface: ChatGPT Canvas or Claude Artifacts land edits in place rather than a full rewrite.

9. Targeted Edit

When to use: a document is mostly right and you want one change, not a fresh draft.

Change only [the specific thing].
Keep everything else exactly as is.
Show me the changed part, not the whole document.

10. Critique-then-Revise

When to use: quality matters and you want a review step before the rewrite.

First, critique this [piece] against: [clarity / accuracy / tone].
List the top 3 weaknesses.
Then rewrite it fixing those, and tell me what you changed.

11. Tighten-and-Trim

When to use: it's too long and you need to cut without losing meaning.

Cut this to [N words / half its length].
Keep every key point. Remove filler and repetition.
Don't add anything new.

Research Patterns

12. Cited Research

When to use: you need current, verifiable facts, not the model's recollection.

Research [topic] using current sources.
For each claim, link the source.
Flag anything you couldn't verify.
End with a 3-bullet summary of what's well-established vs. disputed.

Turn on web access first: browsing in ChatGPT, or Gemini for live Google results. Always click through to confirm links exist.

13. Devil's Advocate

When to use: pressure-testing a plan, decision, or belief before you act.

I'm planning to [decision/plan].
Argue the strongest case against it.
List the 3 most likely ways this fails and an early warning sign for each.
Then tell me whether the plan still holds.

Workflow Patterns

14. Plan-then-Execute

When to use: a multi-step task where a wrong turn early wastes the whole run.

Task: [the goal].
First, give me a numbered plan. Do not start yet.
After I approve, do step 1 and stop for my check.
Then continue one step at a time.

It is the safest way to run a Code Interpreter data job or build a long Artifact — catch mistakes at step one, not step ten.

15. Meta-Prompt

When to use: you're unsure how to phrase the request, so ask the model to write it with you.

I want to achieve [goal] but I'm unsure how to ask.
Interview me: ask the 5 questions you most need answered.
Then write the ideal prompt I should send, and explain your choices.

This is also the fastest way to author instructions for a GPTs build or a Claude Projects custom instruction.

Combining Patterns

Patterns stack. A research-to-deliverable run chains four: Cited Research (12) to gather material, Extract-to-Table (5) to structure it, Draft-and-Outline (6) to build it, and Critique-then-Revise (10) to polish. Keep the steps in one conversation so context carries, and wrap them in Plan-then-Execute (14) if the chain runs long.

Related References

• AI Tool Power-User Rules
• Platform Features Cheat Sheet
• Troubleshooting & Common Issues
• New Feature Evaluation