Better AI habits. Fewer interrupted sessions.

Quick wins

• Use Projects for recurring work.
• Put stable instructions in project instructions instead of pasting them every time.
• Upload reusable files once instead of attaching them again and again.
• Start a new chat when the topic changes.
• Use faster models for simple tasks and stronger models for deep analysis.
• Ask for a compact summary before a conversation becomes too long.

Best Claude workflow

Use Claude Projects as your "context home." If you work on the same client, product, document set or research topic every week, do not rebuild the context from scratch.

A good Claude Project should include:

• The goal of the project.
• The tone or output format you expect.
• The few reference files Claude really needs.
• A short explanation of what should not be repeated.
• A clear instruction to ask before producing very long outputs.

Copy-paste prompt

Before answering, use only the context that is still relevant to this task.
Ignore old details from this conversation if they are not needed.
Give me the shortest useful answer first, then ask if I want a deeper version.

When to start a new Claude chat

Start fresh when:

• You changed client, task or goal.
• The conversation contains old decisions that are no longer true.
• Claude starts referencing outdated context.
• You are asking for a simple task after a long research session.
• The answer feels slower, less focused or too influenced by earlier messages.

Quick wins

• Use /usage to see what the session is consuming.
• Use /clear when switching tasks.
• Use /compact before a long session becomes messy.
• Keep CLAUDE.md short and focused.
• Move rare procedures into skills instead of loading them every time.
• Disable MCP servers you do not need for the current task.
• Avoid dumping huge logs unless the agent really needs them.
• Use deeper thinking only for hard architecture, debugging or planning.

Keep CLAUDE.md lean

Your CLAUDE.md should not be a full internal wiki. It should be a map.

Good CLAUDE.md content:

• How to run tests.
• How to lint or typecheck.
• Project structure overview.
• Naming conventions.
• Critical safety rules.
• Where to find deeper docs.

Bad CLAUDE.md content:

• Full product specs.
• Long historical decisions.
• Every edge case ever found.
• Huge examples.
• Repeated instructions that only apply to one task.

Copy-paste prompt

Before editing, inspect only the files needed for this task.
Do not scan the whole repository unless you can explain why.
Keep test output compact and summarize failures instead of pasting long logs.

Use /clear aggressively

The easiest Claude Code mistake is continuing a new task inside an old session. If you just finished a bug fix and now want to refactor a different area, clear the session first.

Use this pattern:

1. Finish the current task.
2. Ask Claude Code for a final summary.
3. Save anything important in the repo or notes.
4. Run /clear.
5. Start the next task with a tight prompt.

Quick wins

• Use Projects for recurring work.
• Use custom instructions for stable preferences.
• Use temporary chats for one-off tasks you do not want in memory.
• Keep Thinking models for hard problems.
• Use faster/default models for simple drafting, formatting and ideation.
• Ask for "answer first, reasoning only if needed."
• Split heavy research from final writing.

Best ChatGPT workflow

Use three layers:

• Custom instructions for your global preferences. Example: tone, formatting, language, how concise you want answers to be.
• Projects for recurring work. Example: one project for your startup, one for a client, one for SEO research, one for coding.
• Fresh chats for isolated tasks. Example: rewrite a paragraph, generate ten title ideas, clean a CSV column, summarize one short document.

Copy-paste prompt

Give me the useful answer directly.
Do not over-explain unless I ask.
If this task does not require deep reasoning, keep the response compact.

When to avoid Thinking mode

Do not use a Thinking model for:

• Short rewrites.
• Simple summaries.
• Title ideas.
• Formatting.
• Translation drafts.
• Basic code snippets.
• Email polishing.

Use Thinking for:

• Complex debugging.
• Strategic decisions.
• Technical architecture.
• Long synthesis.
• Multi-step reasoning.
• Comparing tradeoffs.

Quick wins

• Keep AGENTS.md short.
• Scope the task to one feature, bug or folder.
• Tell Codex which files matter.
• Avoid asking it to "inspect the whole repo" unless needed.
• Limit MCP servers and external tools.
• Use smaller models for routine tasks when available.
• Ask for compact diffs and short explanations.
• Separate planning from execution.

Good Codex prompts are scoped

Weak prompt:

Improve this app.

Better prompt:

In the billing settings page, fix the empty state when a user has no invoices.
Only inspect the billing route, invoice components and related tests unless you find a direct dependency.
Return a short summary and the files changed.

Keep AGENTS.md useful, not huge

A strong AGENTS.md should answer:

• How does this repo work?
• How should Codex run checks?
• What files or folders should it avoid?
• What conventions matter most?
• What counts as done?

It should not include every product detail, every process document or every example you have ever written.

Copy-paste prompt

Keep the context narrow.
Before reading extra files, explain why they are needed.
Prefer a small targeted change over a broad refactor.
Summarize long outputs instead of repeating them.

Quick wins

• Use the smallest context needed.
• Do not upload huge files when a relevant excerpt is enough.
• Use Fast models for routine work.
• Save Pro or Thinking modes for high-value reasoning.
• Restart or summarize very long chats.
• Use cached context for repeated API workflows.
• Reduce image or video resolution when high detail is not needed.

Large context is not free discipline

A large context window is useful when the model truly needs the whole document set. But many tasks only require:

• The relevant section.
• A summary of the source.
• A table of key facts.
• A few examples.
• A narrow instruction.

Before sending a huge file, ask yourself: "Does Gemini need all of this to answer well?"

Copy-paste prompt

Use only the parts of the uploaded material that are relevant to this question.
If the full file is not needed, ignore the rest.
Give me a compact answer and list any missing context only if it blocks the task.

Multimodal usage tip

Images, video and audio can consume a lot of tokens. If your task does not require small visual details, use lower resolution, shorter clips or a screenshot of the exact part you want analyzed.