01 / service

Find out what your AI is actually reading

Every call carries context. Most of it doesn't apply to the question you're asking. You pay for it anyway, in tokens, in latency and in worse answers.

An audit is a read of your real setup, not a questionnaire. We go through the instructions, prompts, retrieved documents and tool definitions that go out on a typical call. Then we work out which parts earn their place.

Then you get a written note: what's costing you, what's making answers inconsistent and the order we'd fix it in. You own the note whether or not you work with us again.

Some of what we find, you can fix yourself in an afternoon. We've put those below, because a service that depends on you not knowing them is not much of a service.

What a call is actually made of

This is one call. Every harness you use is doing this underneath, once for every step it takes rather than once per question. 5 arguments, and you resend 3 of them in full every single time.

one call, sent again on every turn

client = anthropic.Anthropic()message = client.messages.create(    model="claude-opus-5",  [1]    max_tokens=1024,  [2]    system="You are a support agent. Rules: ...",  [3]    tools=[{...}, {...}, {...}],  [4]    messages=[  [5]        {"role": "user", "content": "..."},        {"role": "assistant", "content": "..."},        {"role": "user", "content": "Where is my order?"},    ],)

Same 5 arguments whether you call the API directly, through an SDK or through a harness like Claude Code. The names change. The bill works the same way.

1model: Which model you are using. It sets the price per token and the size of the window everything else has to fit inside.
2max_tokens: The ceiling on the reply. Output costs several times more per token than input on every current model, so this is the one argument that caps the expensive half of the bill. Set it too high and nothing breaks, which is why it is rarely revisited.
3system: Your standing instructions: rules, tone, project context. Sent in full on every call, whether the question needs them or not. This is the file that quietly grows until it is the largest thing in the request.
4tools: The full definition of every tool the model is allowed to call. Names, parameters, descriptions. All of them go out on every call, including the ones this question will never touch.
5messages: The conversation so far. The API keeps no memory between calls, so the entire array is resent each turn. A long session is not expensive because it is long. It is expensive because it is resent.

Things you can do this week, without us

Set your own compaction point with /autocompact 256k. Compaction already runs by default, so this moves the line rather than adding it: compact earlier and each call carries less, at the cost of losing history sooner. Needs Claude Code 2.1.221 or newer.
Split the one big instructions file. Keep the root CLAUDE.md lean, push folder-specific rules into the folders they apply to and move occasional workflows into skills that load only when something calls them.
Audit your own hooks, skills and commands. A skill costs you its name and description at session start. Its body only loads when something calls it. Hooks and always-on instructions load in full, and you pay for those on every call after.
Trim tool output before it goes back into the conversation. A full API response where 4 fields were needed is the most common quiet cost we find.

What you get

A breakdown of what loads on a typical call and what it costs you
The specific context that can be moved, tiered or dropped
Where retrieval is pulling in noise instead of signal
A restructure plan, ordered by what saves the most for the least work
A memory layer where one is worth having: a linked vault in Obsidian for decisions and history, or a vector store like Pinecone for the bulk, so your own data is reachable on request instead of loaded on every call
Rough before and after figures, with the baseline stated
A walkthrough of the findings with your team, or the changes made by us and training on what we changed

This fits if

You already run AI on something that matters
Your bill or your latency is growing faster than your usage
Answers are inconsistent and nobody can say why
You have one big instructions or system prompt file that keeps growing

This isn't for you if

You haven't shipped anything with AI yet. There's nothing to audit, so start with training instead. We run courses at every level, from a team that has never used it to one already building agents.
You want someone to pick a model for you. That is a shorter conversation.
You want the report to say the tool you already bought was the right call.

How an audit runs

A short call

You show us what you run and what hurts. 30 minutes, no preparation needed.

We read the setup

Read-only access to the prompts, config and a sample of real calls. We don't need production credentials.

You get the review

Written findings and a fix order. We walk your team through it, or we make the changes ourselves and train your people on what changed. Either way you decide.

Questions about audits

What access do you need?

Read access to your prompts and config, plus a sample of real calls with anything sensitive stripped. We don't need write access or production credentials.

How long does it take?

Most audits take about a week from the first call to the written note. Bigger setups take longer, and we tell you before we start.

Do you implement the fixes?

Your call. We write the review so your own developers can act on it, and plenty of clients do exactly that. If you would rather we made the changes, we do that and then train your team on what changed, so the knowledge stays with you.

What if the savings are small?

Then the note says so. We'd rather tell you your setup is already tight than invent a problem.

Curious what a call actually costs you?

Show us the setup. We'll tell you within 30 minutes whether an audit is worth your money.

Talk to us about your setup See the other services