AI coding cost audit checklist

A practical checklist for finding coding-agent token waste.

Use this checklist to review Claude Code, Cursor, Codex, Copilot, MCP servers, custom agents, prompt workflows, memory, and model routing. It is designed for teams that know usage is rising but do not yet know where the waste is.

Request audit review View full audit

Usage baseline

List the coding agents, IDE agents, CLI agents, MCP servers, and custom tools currently in use.
Identify the teams, repos, and workflows where usage is highest.
Separate experimentation, normal coding, migration work, review work, and batch agent runs.

Repeated context

Find instructions engineers paste or recreate at the start of sessions.
Identify repo facts, commands, architecture notes, and debugging history agents repeatedly rediscover.
Mark what should become repo instructions, retrieved memory, or reusable task briefs.

MCP and tools

Count which tools are always exposed to agents before a task starts.
Review heavy schemas, long descriptions, examples, enums, and verbose responses.
Split always-on tools from task-specific tools where possible.

Prompt and workflow reuse

Collect prompts and task formats that consistently produce good results.
Turn repeated workflows into commands, templates, or playbooks.
Define review steps for high-risk agent output.

Model routing

Identify tasks that need frontier reasoning versus tasks that need retrieval, scripts, or smaller models.
Move deterministic transformations and lookups out of expensive chat loops.
Add eval cases before changing routing for important workflows.

When to ask for help

The checklist is useful when it turns into specific questions.

If the review surfaces repeated context, unclear model routing, heavy MCP schemas, missing memory, or agent workflows that are hard to evaluate, the audit can turn those symptoms into a scoped implementation plan.

Start diagnostic