What You Need to Know
Overhead is anything you type, explain or attach at the start of every similar conversation. Optimising means making your work with Claude cost less time or usage. A measurement times the overhead and counts how often it recurs. This objective covers how you work with Claude, not the business process around it (Domain 4.2).
A kitchen fitter drafts customer quotes with Claude. Each quote conversation opens with how his business works and prices a job, plus the same price list and catalogue. Timed, that is about ten minutes, three or four times a week.
Measure First
The difference between optimising and fiddling is a number. His measured ten minutes justify moving the preamble somewhere permanent. "Quoting feels clunky" justifies nothing yet except timing it. This is Domain 4.5's honesty rule pointed at your own workflow: improve what you have observed, and claim no saving you did not measure.
The Repeated-Overhead Test
Within a single conversation, Claude remembers the context from earlier messages. The fitter's preamble recurs because he types it again at the start of each new quote conversation. Recurring overhead belongs somewhere it only has to exist once.
Standing context and rules, such as his pricing rules, go in a Project's instructions. Anthropic describes these as tailoring Claude's responses for each project. They suit work that needs Claude to keep consistent context across many conversations in the same Project. Recurring reference material, such as his price list, goes in Project knowledge, making the Project a self-contained workspace. A request made in nearly the same words every week becomes a saved form you adapt per use. Preferences that apply to everything you do belong in the account-wide personalisation settings, where they apply to all conversations.
Batch Related Requests, and Name Your Files
Two documented levers cost minutes. First, batching. Anthropic's usage guidance says to group multiple related tasks or questions in a single message rather than fragmenting them across exchanges. It also says to be specific and concise, because vague queries can leave Claude needing clarification. The fitter asks for the quote and cover email in one message. Anthropic's article on usage and length limits adds that longer conversations which trigger automatic context management consume more of your usage limit. So letting working chats sprawl costs usage as well as tidiness.
Second, file names. In a large Project, retrieval decides which material Claude works from. Anthropic states that well-named files help Claude understand and retrieve the right information, alongside referencing documents by name and grouping related documents in the same Project. His "pl_v3_final" becomes "Worktop price list". Renaming a file so a stranger could tell what it contains is a real optimisation with a documented mechanism.
Match the Settings to the Work
Running everything at one setting wastes in both directions. Higher effort produces more thorough responses but takes longer and uses more of your usage allowance. Routine reformatting at maximum effort pays for depth it never uses. Deep analysis rushed at a low setting underserves the task. A pattern of slow, over-worked responses to routine work, such as his quote reformatting, is itself the measurement pointing at a mismatched setting.
The Proportion Rule
Some overhead is not worth removing. A twenty-minute task the fitter does once a year does not earn a Project, a saved form or an afternoon of setup. The arithmetic never pays back, and the maintained artefact goes stale between uses. Recognising the optimisation that costs more than it saves is as professional as spotting the one that pays.
In the exam
Most items turn on whether the option in front of you has a number. What is tested is noticing, from a symptom or a measurement, where an optimisation belongs, including which piece of a repeated preamble goes at which level. Sometimes it belongs nowhere: in this domain's hardest items, deciding not to optimise, on the numbers, is keyed. Choosing model and effort is Domain 3.3's decision, and writing the configuration is Domain 5's work.
Key Concept
In 7.3 stems, look for the measurement. The keyed answer moves measured, recurring overhead to where it exists once, or declines a setup that costs more than it saves. Distractors optimise something untimed, such as an unmeasured reorganisation, or everything indiscriminately.
Exam Traps
Sweeping workflow changes proposed off a feeling
Without a measurement there is no baseline, no way to know whether the change helped, and no honest claim to have saved anything. The keyed answer times the work first; the reorganisation can wait for the number.
Building a Project or saved form for a task done twice a year
Setup and upkeep that exceed the overhead they remove are a cost dressed as an improvement, and the artefact goes stale between uses. The proportion rule says leave the rare task alone.
Raising effort or switching model to speed up a workflow whose overhead is a repeated preamble
No setting removes ten minutes of retyped context; that overhead moves into a Project’s instructions and knowledge or it stays. Higher effort also takes longer and uses more of your usage allowance, optimising in exactly the wrong direction.
Practice Scenario
A city farm’s development officer starts every funder conversation the same way: explaining what the farm is and which programmes it runs, then attaching the same three documents. Timed, that comes to about ten minutes, three or four times a week. Each January she also spends twenty minutes preparing the farm’s annual return for the charity regulator. She will put effort only into what her timings show is costing time, and only where the saving outweighs the set-up. What is the most appropriate optimisation?
Build Exercise
Time the Overhead, Then Place It or Leave It
What you'll learn
- How to measure repeated overhead honestly
- How to place each piece of overhead at the level where it exists once
- How to use file names as a retrieval lever
- How to apply the proportion rule and decline an optimisation
- For one week, note what you type, explain or attach at the start of every similar Claude conversation, and time it honestly
Why: Optimisation starts from a measurement; the week of notes turns a feeling of clunkiness into numbers you can act on.
You should see: A short list of recurring overhead items, each with a rough time cost and how often it recurs.
- Apply the repeated-overhead test to each item: standing context belongs in a Project’s instructions, recurring reference material in its knowledge, and a near-identical weekly request in a saved form you adapt
Why: Each kind of overhead has a level where it only has to exist once, and placing it wrongly means either repeating it anyway or burying it where it does not apply.
You should see: Each overhead item labelled with its destination: instructions, knowledge, saved form, or account-wide preferences.
- Check the names of any files already in your Projects, and rename anything a stranger could not identify from its filename alone
Why: Retrieval decides which material Claude works from in a large Project, and well-named files are a documented input to getting the right one.
You should see: File names that state what each document contains, replacing versions, initials and abbreviations only you understand.
- Find one repeated task you run at a mismatched setting: routine work at maximum effort, or deep analysis rushed at a low setting, and match the setting to the work
Why: Settings mismatched to the work waste time and usage in one direction and quality in the other; a pattern of over-worked routine responses is itself a measurement.
You should see: One task moved to a setting that fits it, with a note of what changed in speed or depth.
- Pick your rarest repeated task and do the proportion arithmetic: if the setup would cost more than a year of the overhead it removes, leave it alone and note that you decided to
Why: Declining an optimisation on the numbers is the discipline that separates professional optimisation from tooling up for its own sake.
You should see: A one-line record of the task, the arithmetic, and the deliberate decision not to optimise it.
Sources
- What are projects? — Anthropic
- Understanding Claude's personalization features — Anthropic
- Usage limit best practices — Anthropic
- How do usage and length limits work? — Anthropic
- Retrieval augmented generation (RAG) for projects — Anthropic
- Change the model, effort, and thinking settings — Anthropic
- Claude Certified Associate – Foundations Exam Guide — Anthropic. The measure-first and proportion rules draw on the exam guide's objective and standard practice. The batching, file-naming, instruction and settings levers come from the Help Center articles above.