Stop burning your best model on busywork

A Claude Code cost saving workflow that keeps the quality where it matters. Your best model plans and reviews, a cheaper one does the typing, and the loop keeps going until the job is actually done.

Most people run one model for everything: the plan, the edits, the formatting, the test fixes. Then the usage limit lands somewhere around lunch. The fix is not a cheaper model, it is a division of labour.

Claude Code is Anthropic’s coding agent. It reads files, runs commands and works through multi-step jobs with very little handholding. It can also put different models on different parts of one job, which is the part most people never switch on. On Anthropic’s published API rates, Haiku 4.5 costs 1 dollar per million input tokens against 5 for Opus 5 and 10 for Fable 5, so where the volume sits decides most of your bill.

Nothing extra needs installing. Everything below is a setting or a command inside Claude Code, and /goal needs version 2.1.139 or later, so run claude --version if you are unsure.

01 / The pattern

The Claude Code cost saving workflow in five steps

Your expensive model touches the work twice, at the start and the end, where thinking matters. The cheap one carries the middle. That split is the whole saving.

01

Let your best model plan, not build

Start in your premium model and ask for a plan rather than code: the steps, the files it will touch, and a clear test for “done”.

why it savesNo good team lets its most expensive person do the typing. Planning is short; building is long, and it needs no judgement once the plan is agreed.

plan mode →

plan prompt, fill in the blanks

Before writing any code, give me a step-by-step plan to [what you want built or fixed] in [which files or part of the app]. List the files you will change, the order, and a checkable test for “done”, for example tests passing or a specific result. Do not build yet. Wait for my go-ahead.

There is a one-command version of this too. Set your model to opusplan and Claude Code plans on Opus, then drops to Sonnet automatically to build.

02

Hand the grunt work to Haiku

Once the plan is approved, make a small helper called a subagent, point it at Haiku, and let it carry out the approved steps: edits, boilerplate, tests, formatting.

why it savesA subagent is just a text file with a model setting, so the repetitive middle of the job moves to the cheapest model without changing how you work.

subagent docs →

03

Bring your best model back to review

Switch back to the premium model, or use a reviewer subagent set to it, and have it critique what the builder produced. Claude Code ships with a /code-review skill you can use for this.

why it savesThe builder is a poor judge of its own work: same context, same blind spots. A reviewer that did not write the code catches far more, and review is cheap because it is short.

bundled skills →

If you have a house style, tests or known edge cases, write them down once as a SKILL.md file. A skill is a saved set of instructions Claude loads when it needs them, so the check is the same every run rather than whatever you remembered to mention.

04

Set the bar with /goal so it will not stop early

Give Claude a condition it has to satisfy before it is allowed to stop, plus a cap on how many turns it may take getting there.

why it savesLeft alone, an agent stops the moment it thinks it is done, which is usually too soon. With /goal a separate small model checks your condition after every turn and sends it back until the bar is met.

goal docs →

05

Let it run on its own with /loop or /schedule

/loop re-runs a prompt on a timer inside your open session. /schedule saves it as a routine, a job that runs on Anthropic’s cloud, so it keeps working with your laptop shut.

why it savesRepeating work stops needing you at the keyboard. The trade is that an unattended loop spends tokens whether the run was useful or not.

routines docs →

Written out, the three commands look like this:

• /goal get all tests in /tests passing and no new console errors, stop after 5 tries sets a finish line the checker can verify, and a cap on how long it may chase it.
• /loop 15m check my PR, address review comments, fix failing CI runs on your machine every fifteen minutes and stops when you close the session.
• /schedule every hour: triage new bug reports and draft a fix saves the same prompt as a cloud routine that survives a closed laptop.

before you leave one running

Routines, which is what /schedule now creates, are still a research preview, so limits and behaviour can change without notice. An unattended loop also bills you for every run, useful or not. Watch the first few, then match the interval to how often the inputs actually change.

02 / The build agent

Setting up the builder subagent

A subagent is a small helper with its own instructions and its own model. Ours is called builder, it runs on Haiku, and its only job is to carry out plan steps that have already been approved.

06

Ask Claude Code to write it for you

Type this into a session: “Make a subagent called builder that runs on Haiku and carries out approved plan steps.” It creates the file, in the right folder, without you touching anything.

why it savesYou never have to find a hidden folder or get the file format right. The rest of this section is only if you want to see what it made, or prefer to set it up by hand.

subagent docs →

the file it creates: builder.md

--- name: builder description: Runs approved steps, edits, tests, formatting. model: haiku ---

Follow the plan exactly. Run tests. Flag anything unclear.

07

Know where it lives

The file sits two folders deep at the top of the project you are working in, at .claude/agents/builder.md.

why it savesKept in the project, it applies to that project only. The same folder in your home directory, at ~/.claude/agents/, applies to every project you open.

subagent docs →

That path looks like a scary filename, but it is not one. It is a trail through folders, where each slash means “go inside”:

• your-project/ the folder you are building in
• .claude/ a folder inside it, hidden because the name starts with a dot
• agents/ a folder inside that
• builder.md the only real file, plain text, where .md just means Markdown

On a Mac, press Cmd + Shift + . in Finder to see the hidden .claude folder. And if you edit the file by hand, restart Claude Code so it loads the change.

03 / Guardrails

Three rules that keep a loop honest

The savings are real, but a loop is only ever as good as its stop condition and its reviewer. A vague goal plus a cheap model with nobody checking gives you fast, cheap and wrong, at scale.

08

Judgement stays on the strong model

Route the decisions to your premium model: the plan, the review, and anything ambiguous. Haiku executes, it does not decide.

why it savesCheap tokens spent going the wrong way are not cheap. The saving comes from moving volume, not from moving judgement.

model settings →

09

Every loop needs a real check

Close it on something Claude cannot talk its way past: tests passing, a build exiting cleanly, a score threshold, an empty queue. Then cap the turns.

why it savesNo cap means no ceiling on spend. Conditions that are either true or not beat conditions like “the code is good”, which a model can always argue it has met.

goal docs →

10

Watch the meter

Run /usage to see where your usage actually went, broken down by skill, subagent, plugin and connector, with a toggle for the last day or the last week.

why it savesGuessing where the tokens go is how people end up rationing the wrong thing. Pilot on a small slice before any big or scheduled run.

cost docs →

Watch out

Common mistakes when automating Claude Code in 2026

The setup is the easy part. Most of the money gets lost afterwards, in ways that look reasonable at the time:

• Leaving the builder on inherit. If its model line does not say haiku, the subagent quietly runs on whatever your main session is using, and you save nothing.
• Skipping the description line. It is required, and it is the sentence Claude reads to decide when to call the agent. Leave it out and the agent just sits there, never used.
• Looping faster than the inputs change. Checking a pull request every five minutes when reviews arrive twice a day is paying to watch a static page.
• Walking away on run one. Watch the first few runs of anything unattended. That is when a stop condition turns out to be wrong.
• Using loops for open-ended creative work. They suit well-defined, repeating jobs like triage, CI fixes and migrations. Where “good” is a matter of taste, your judgement still belongs in the loop.

Questions

FAQ

Will this actually cut my costs?

It cuts them if most of your tokens were going on execution, which for coding work they usually are. Planning and review are short; editing files and re-running tests is long, and that is the part that moves to the cheaper model. Check /usage before and after rather than trusting any headline percentage.

Do I need to be a developer to set this up?

You need to be comfortable running Claude Code, but you do not need to write any files yourself. Ask it in plain English to make a subagent called builder that runs on Haiku, and it writes the file in the right place for you.

Is Haiku good enough for the building step?

For well-specified work, usually yes. It is fast, it is cheap, and it follows a clear plan reliably. The quality here comes from the plan being good and the review being done on a stronger model, not from the builder being clever.

What is the difference between /goal and /loop?

/goal decides when to stop: after every turn a separate small model checks your condition and sends Claude back if it is not met. /loop decides when to start again, on a timer, and keeps going until you stop it. Use /goal to finish something, /loop to watch something.

Does /schedule keep running when my laptop is closed?

Yes. /schedule creates a routine that runs on Anthropic’s cloud rather than on your machine, so a closed laptop does not stop it. Routines are still a research preview, so expect the limits and the interface to shift.