Claude Efficiency Kit

Double your Claude
usage. Keep everything.

By Dan Donovan, @askdandonovan

The model cheat sheet, CLAUDE.md template, and five copy-paste prompt templates that stop the most common ways people burn through their Claude limit.

15 Tips covered
5 Prompt templates
1 CLAUDE.md file
$0 Extra cost
What's inside
Model Cheat Sheet
Sonnet 4.6 vs Opus 4.7: when to use each and what the wrong choice costs
CLAUDE.md Template
Five lines that make Claude Code responses noticeably shorter and more direct
5 Prompt Templates
Copy, fill in the blanks, send. Get the right output first try every time
Hidden Cost Breakdown
Images, adaptive thinking, long responses: what's silently burning your budget
15-Tip Checklist
Every tip from the full series (tap to check off as you implement each one)
Prompts Library Framework
How to build your own saved prompts so you never start from zero again
Start here
Do these 3 things right now. Everything else is a bonus.
1

The model decision

The single biggest lever on your Claude usage is which model you're sending messages to. Most people default to Opus for everything. That's the most expensive mistake you can make.

Sonnet 4.6 Opus 4.7
Token cost per message Base cost ~1.7× more per token
Adaptive thinking Not supported Supported (when enabled, adds reasoning tokens you're billed for but never see)
Image token cost Standard Up to 3× more
Best for 90% of everyday tasks Complex strategy, hard debugging, creative work where Sonnet isn't enough
Avoid for Tasks where quality genuinely requires deeper reasoning Reformatting, summarising, quick Q&A, simple rewrites
The Opus 4.7 hidden cost: Opus 4.7 supports adaptive thinking. When enabled, Claude reasons through your request before responding and you're billed for every token of that reasoning, even though you never see it. Combined with the higher base price, Opus can cost significantly more per task than the 1.7× headline ratio suggests.

The practical rule: start every task on Sonnet 4.6. Switch to Opus only if the output genuinely isn't good enough. Most tasks won't require the switch.

2

The CLAUDE.md template

CLAUDE.md is a file you create in your project folder. Claude Code reads it automatically before every response. Five lines in this file make Claude Code responses noticeably shorter and more direct, without changing the quality of the output.

Create a file called CLAUDE.md in your project folder and paste this:

CLAUDE.md 
# Response Rules
- No padding or filler sentences
- Short sentences — one idea at a time
- Direct answers only — skip restating the question
- Start with the answer, not a preamble
- No openers like "Great question" or "Sure, I can help with that"
Claude Code only. This file is read by Claude Code, not the Claude.ai web interface. If you use Claude.ai, Section 3 covers how to get the same result using a prompt instead.
3

5 prompt templates

These work in Claude.ai, Claude Code, and any Claude interface. Copy each one, fill in the parts in brackets, and send. Each is designed to get the right output on the first try, no back-and-forth.

Context Handoff
Use when you need to switch to a new chat mid-task
Summarize this conversation into bullet points for a new chat.
Include: key decisions made, important context, open questions, next steps.
Be concise — this is going into a new chat as context, not a document.
Example output
"We're writing a proposal for a 20-person agency. Decided: 3-page format, fixed fee not retainer. Next step: draft the pricing section. Open question: how to frame the guarantee clause."
Format Spec Upfront
Use to prevent long exploratory responses on simple requests
[YOUR QUESTION OR TASK]

Format: [bullet points / numbered steps / one paragraph / table]
Length: [under 100 words / 3–5 bullets / one short paragraph]
Output only — no commentary before or after.
Filled example
"What's the difference between Claude Projects and the Claude profile setting? Format: bullet points. Length: under 80 words. Output only."
Edit, Don't Rewrite
Use when you want targeted changes, not a full rework
Edit this [email / document / post] for [GOAL: clarity / shorter / stronger CTA].
Change only what's necessary to hit that goal.
Don't rewrite sections I didn't ask you to change.
Return the full revised version.
Filled example
"Edit this client email for shorter and more direct. Change only what's necessary. Don't rewrite sections I didn't ask you to change. Return the full revised version."
One Thing Only
Use instead of open-ended "what do you think?" questions
Look at [THING YOU'RE WORKING ON].
Don't give me a full breakdown.
Tell me: what is the single most important change I should make, and why?
One recommendation only.
Filled example
"Look at this landing page headline: [paste headline]. Don't give me a full breakdown. Tell me: what is the single most important change I should make, and why? One recommendation only."
First Draft Fast
Use when starting any written piece from scratch
Write a [TYPE: email / post / proposal / summary] for [AUDIENCE] about [TOPIC].
Goal: [what you want the reader to do or feel]
Tone: [direct / conversational / formal]
Length: [under X words]
Output only — no preamble.
Filled example
"Write a LinkedIn post for small business owners about switching from Opus to Sonnet 4.6 to save Claude usage. Goal: get them to try the switch today. Tone: direct. Length: under 150 words. Output only."
4

Hidden costs most people never catch

These are the things burning your Claude budget that nobody covers.

~1,200
Words of text a single 1080p screenshot costs in tokens
3×
More expensive per image on Opus 4.7 vs Sonnet 4.6 (more tokens + higher per-token price)
Hidden
Adaptive thinking tokens on Opus 4.7, billed but never shown in the response
📸
Images cost as much as a full page of text
A standard 1080p screenshot costs roughly 1,568 tokens, about 1,200 words of plain text. If you can describe what's in the image in two sentences, write the two sentences instead of screenshotting it.
🧠
Opus 4.7 adaptive thinking adds hidden token costs
Opus 4.7 supports adaptive thinking. When enabled, Claude reasons through your request before responding. You're billed for every token of that reasoning, but none of it appears in the response. Combined with Opus's higher base price (1.7× more than Sonnet), the total cost per task can be significantly higher than expected.
📄
Long Claude responses stay in context and keep costing you
If Claude wrote a 2,000-word document earlier in your chat, it re-reads all 2,000 words before answering your next message. Once you have the output you needed, start a new chat rather than keep that response in play.
🔄
Every refinement round re-sends your full history
Going back and forth to get the output right re-sends the entire conversation each time. Three vague prompts and two corrections cost far more than one well-specified prompt. Use the Format Spec template above.
5

The full 15-tip checklist

Every tip from the series. Tap any tip to check it off as you implement it.

The five foundations
  • Quick · 10 secSwitch your default model to Sonnet 4.6. Use Opus only when Sonnet genuinely can't do the job.
  • Quick · 30 secStart a new chat every time you switch tasks.
  • Quick · 2 minWhen you can't start fresh, ask Claude to summarise the conversation into bullet points and paste that into a new chat.
  • Quick · 30 secOnly upload the pages or sections Claude actually needs to read, not the full document.
  • Setup · 3 minSet up your Claude profile once in Settings › General, under personal preferences. Every new chat starts with that context already loaded.
Hidden costs to eliminate
  • Quick · 30 secPaste text instead of screenshots wherever possible. Images cost ~1,200 words of tokens each.
  • Quick · 30 secStart a new chat after Claude produces a long document. Don't keep long responses in context if you're done with them.
  • Quick · 10 secAvoid Opus 4.7 for simple tasks. It costs 1.7× more per token than Sonnet 4.6, and adaptive thinking adds further hidden costs when enabled.
  • Quick · 30 secSpecify your output format and length before asking the question. Eliminate refinement rounds.
  • Quick · 10 secDon't ask Claude to confirm it understood before giving it the task. Just give it the task.
Habits that compound over time
  • Quick · 2 minWhen an ongoing task needs to continue, summarise the conversation first, then paste into a new chat.
  • Setup · ongoingSave every prompt that gets you the right output on the first try. Build a library over time.
  • Quick · 10 secTell Claude how long you want the answer before asking. It defaults to comprehensive responses.
  • Setup · 5 minUse Claude Projects for recurring work. Set the context once; every chat in that project starts with it loaded.
  • Quick · 10 secAsk specific questions, not open-ended ones. "What's the one thing I should change?" not "What do you think?"
Building your prompts library: Every time you write a prompt that gets the right output on the first try, save it. A Notes file, Notion page, or simple text file works fine. Within a month you'll have 20 to 30 prompts covering most of your recurring tasks. You'll stop burning tokens on back-and-forth and stop starting from scratch every time.