Rules & Best Practices

These are the durable habits that separate casual users from power users of Claude, ChatGPT, and Gemini. Treat each heading as a rule you can apply in any app. For platform specifics, see the Cheat Sheet; for ready-made structures, see Prompting Patterns.

Conversation Hygiene

How you manage individual chats matters more than which model you pick.

One topic per conversation

Start a fresh chat when you switch subjects. Old, unrelated turns crowd the context window and pull answers off-target, and a focused chat is easier to find later.

Restart instead of fighting

If a conversation has gone wrong three times, do not keep correcting it — the bad attempts stay in context and bias every reply. Open a new chat and paste a clean, improved request.

Edit and branch rather than pile on

Every desktop app lets you edit an earlier message and regenerate from there, creating a clean branch. Use it instead of a long "actually, ignore that" correction.

Name and organise your chats

Rename important conversations and pin the ones you return to. In ChatGPT, hover a chat in the left sidebar for the three-dot menu; in Claude, click the chat title to rename it.

Context Management

The context window is finite working memory. Treat it as a scarce resource.

Front-load the important stuff

Put role, goal, constraints, and key facts at the top of your message. Models weight early and recent text most heavily, so requirements buried in a wall of text get missed.

Summarise to reset, do not just continue

When a long chat repeats itself or ignores instructions, ask it to summarise the decisions so far, then paste that summary into a new chat. You keep the conclusions and drop the clutter.

Summarise this conversation as a handoff brief:
- The goal we agreed on
- Decisions made and why
- Open questions and any constraints I gave you
Keep it under 200 words.

Use persistent context for anything recurring

Do not re-explain who you are every session. Capture standing context once and let the platform reuse it.

Watch for the slow-conversation warning

Very long chats become sluggish and forgetful before they hard-stop. When replies slow or the app warns the conversation is long, summarise and start fresh — see Troubleshooting.

File Preparation

Most disappointing file results come from bad inputs, not bad models.

Give text, not pictures of text

A screenshot of a PDF forces the model to read pixels and invites transcription errors. Upload the real document or paste selectable text when you can.

Trim before you upload

Delete cover pages, boilerplate, and irrelevant appendices first. A focused 5-page extract beats a 200-page report where your answer is one line.

Label what you attach

State what each file is and what you want from it. The model does not know your filenames mean anything.

Attached: Q3-sales.csv (monthly revenue by region).
Compare this quarter to last. Call out the three biggest
changes by region and a likely cause for each.

Match the file to the right tool

Respect privacy before uploading

Strip or redact sensitive data — client names, account numbers, health details — unless you are certain of your account's data settings. When in doubt, anonymise first.

Output Quality

Specify the format you want

Ask for a table, numbered list, email, or one-paragraph summary explicitly. "Compare these" yields prose; "compare these in a table with columns X, Y, Z" yields something usable.

Show one example for anything non-obvious

A single example of your desired tone or structure (a one-shot example) beats three sentences describing it. Paste a sample and say "match this."

Ask for the reasoning when stakes are high

For analysis or recommendations, request the working, not just the verdict. "Walk me through your reasoning, then recommend" surfaces shaky assumptions you can challenge.

Make it edit in place

For documents or code you revise repeatedly, use the live-editing surfaces instead of regenerating whole replies: Artifacts in Claude and Canvas in ChatGPT keep an editable panel beside the chat.

Verify, do not trust

Treat confident-sounding facts, quotes, and statistics as claims to check. For anything time-sensitive, prefer a tool with live browsing and ask it to cite sources you can open.

Iterate, do not accept the first draft

The first answer is a starting point. Name what is wrong specifically — "too formal," "missing the cost angle" — rather than regenerating blindly.

Good, but revise:
1. Tighten the intro to two sentences.
2. Add a row for shipping costs.
3. Keep everything else exactly as is.

The Quick Rules

To judge whether a new feature is worth adopting, see Feature Evaluation.