AI Guides › Playbooks
By Nigel Guy · 8 min read
You open one chat in the morning and use it all day. You send "no, I meant..." corrections, paste in yesterday's brief again, and leave the top model and every tool switched on. Each message feels cheap. Then you hit your limit before lunch and blame the plan, when usually the habits are what's expensive.
The rule: every message you send pays for the whole conversation behind it, so keep that conversation short and clean, and don't send the same thing twice.
Anthropic's help centre lists what uses your allowance: message and conversation length, attachment size, model, effort level, and features such as extended thinking, web search, Research, connectors, artifacts and multi-step tasks. Paid plans run two limits at once: a five-hour session limit and a weekly limit.
The thing most people miss is how conversations are sent. The Claude Code documentation spells it out: the full conversation goes with every request, so "a one-line question in a session that has been open all day still draws usage for the whole conversation." Caching makes re-reading that history cheaper, but it still isn't free. Anthropic doesn't publish exactly how claude.ai counts each message, but it lists conversation length as a main cost.
And claude.ai, Claude Desktop and Claude Code all draw on the same limit, so an afternoon of coding eats your writing allowance.
Six habits, each matched to the cost it cuts.
| # | Habit | What it cuts |
|---|---|---|
| 1 | Edit the original message instead of sending a correction | Failed attempts that every later message has to carry |
| 2 | Batch related asks into one numbered message | A full trip through the history for each small question |
| 3 | One task per chat, with a short handover | A long, stale history on every message |
| 4 | Put reference documents in a Project | Pasting the same material again and again |
| 5 | Match the model and effort to the job | Paying top rate for simple work |
| 6 | Turn off the tools this task doesn't need | Search, Research and connector calls you never asked for |
If a reply misses because your message was unclear, edit that message rather than adding "actually, I meant...". A correction leaves the wrong exchange in the history for every later message to carry; editing replaces it. The edit control sits on your sent message; its exact look varies by app version.
Anthropic's own advice is to put related questions in a single message. Number them: "1. Tighten the intro. 2. Suggest three titles. 3. Flag claims needing a source." One request instead of three.
When the task changes, start a new chat, carrying over only a short handover (prompt below). In Claude Code the same habit is /clear, which the docs say "costs nothing". /compact, on the other hand, has to read the whole conversation to summarise it, so compacting a very large session is a large request in its own right.
Anthropic says documents you upload to a Project are cached, and cached portions "count less against your limits" when you refer to them again while they're still cached. Your style guide and client brief belong in project knowledge, not a fresh paste every morning.
The model picker sits under the message box on web and desktop, and at the top of the screen on mobile. Anthropic's Claude Code guidance applies to chat too. Use the mid-tier model for most work, keep the top model for hard reasoning, and pick a lighter model for simple jobs where your plan offers one. Turn off extended thinking or lower the effort when the task doesn't need deep reasoning. The help centre says higher effort uses more tokens.
Web search, Research and connectors each add work. Rewording a paragraph doesn't need the web, so turn tools on per task, not by default.
The trap is the "handy" mega-chat: one thread for the whole week, opened with a big context dump. Every message then pays for all of it. Two related traps:
/usage. On Pro, Max, Team and Enterprise it breaks down your usage by skills, subagents, plugins and MCP servers, and flags anything such as long context or cache misses that accounts for 10% or more of recent usage. Press d or w to switch between day and week. That breakdown only covers this machine, not claude.ai. /context shows what's filling the current session.Paste this at the end of a long chat. Fill in a focus, or delete that line.
You are reviewing how I've used this conversation, so I can use less of my Claude allowance next time.
Goal: show me specific places where I spent usage I didn't need to. A good answer quotes or points to my actual messages. It doesn't give general tips.
Focus (optional): [ANYTHING_TO_LOOK_AT_FIRST]
Work through this in order:
1. List each of my messages that corrected or clarified an earlier one, where editing the earlier message would have done the job.
2. List groups of my messages that could have been sent as one numbered message.
3. Point out any part of this chat that needed nothing more than a lighter model, lower effort, or no tools.
4. Say whether this conversation should have been split into separate chats, and where the split should have come.
Output: four short headed sections matching the steps. Under each, bullets of the form "Message [n or short quote] → what to do instead". If a section has nothing, write "Nothing found".
Rules: be blunt but accurate. Don't invent messages, and don't count your own replies as mine. If you can't see the earlier part of the conversation (for example, because it has been summarised), say so rather than guessing.
Before you answer, check that every bullet points to a real message of mine.
Fill in the next task, then paste the result into a new chat.
I'm about to start a fresh chat for this next task: [NEXT_TASK]. Write a handover I can paste into it.
Include only what the new chat needs: decisions made, constraints, key facts, and the current state of any draft. Leave out discussion, dead ends and anything unrelated to [NEXT_TASK].
Format: a heading "Handover", then no more than 12 bullets, then one line beginning "Open question:" if anything is unresolved.
If you're not sure whether something is relevant to [NEXT_TASK], ask me before including it. Check that the handover is under 200 words before you reply.
Say you're a freelance copywriter on Pro who hits the session limit around noon. The Send Audit on your morning chat finds four "no, shorter" corrections, three one-line questions about the same email, and a client brief pasted three times. That week you move the brief into a Project, edit instead of correcting, and start a new chat per deliverable with the Handover prompt. You check Settings > Usage at lunch each day. No saving is guaranteed, but you're now measuring rather than guessing.