Mikie / Field notes

Better context.
Better questions.

Ideas, practical workflows, and measured results from the way we work with coding agents.

Select relevant code from a repository, then use it for a focused coding task.REPOSITORYFOCUSED TASK
Less to read. More to work with.
Featured · Benchmark

Mikie vs RTK vs Caveman: the Claude Code cost benchmark

One 72-prompt Claude Code session, the same tasks and the same hidden tests for every tool. Mikie cut the bill 63.7% and passed as many hidden tests as plain Claude Code. Caveman cut it 0.5%. RTK raised it 27.3%.

· 7 min read · Read the story ↗

Ideas into practice

Understand the workflow, see what changes, and try the useful parts in your next session.

Benchmark report12 min read

What Mikie saves in Codex: 493 measured sessions

We ran the same coding tasks through Codex with and without Mikie: same model, same reasoning level, every change checked by hidden tests. On everyday Python changes Mikie cut the cost by 68 percent. Here is everything we measured, including where it falls short.

Explainer6 min read

Fewer tokens is not a smaller bill

Mikie cut tokens 84.6% in long Claude Code sessions. The bill fell 66.3%. Those are different numbers, and the second one is the one you pay. Here is why they differ, and how to check any tool that says it saves you money.

Explainer7 min read

Why your coding agent reads far more code than the task needs

A coding agent can read a large part of your repository to solve a small task. A better starting point helps it focus on the code that matters.

How-to6 min read

Shadow mode: see what Mikie would save before you change anything

Shadow mode runs Mikie alongside Claude Code or Codex without changing what the agent sees. You measure Mikie on your own work first: a labelled projection in Codex, an unchanged session to compare against in Claude Code.

Explainer6 min read

What's on a Mikie receipt, and why your finance team will want one

Every turn your agent takes leaves a receipt: what it used, what it cost, what was saved, and how sure that number is. None of it contains your code.

How-to5 min read

Set up Mikie with Claude Code or Codex

Install the CLI, connect your coding client, and verify one focused task in your own repository.

Field note6 min read

Tokens, credits, and coding limits: what are you actually using?

A token counter, a credit balance, and a subscription limit describe different things. Read the right one before deciding what to optimize.

Field note7 min read

One task, three context choices

A tiny parser fix shows why the right context includes both the change and the check. Here are three ways to frame the same task.

Know the difference

Compare the role each tool plays and choose what fits the way you build.

Closest thesis

Mikie and Augment Code

Augment and Mikie agree that better context beats more context. The difference is where the indexing happens and what you get back afterwards.

Different layer

Mikie and Cursor

Cursor is an AI editor with retrieval built in. Mikie is a context and receipt layer for Claude Code and Codex.

Different layer

Mikie and GitHub Copilot

Copilot is built into GitHub and your editor. Mikie decides what Claude Code or Codex should read, and writes down what each request used.

Different layer

Mikie and Windsurf

Windsurf is an AI editor with its own agent, Cascade. Mikie is a context and receipt layer for Claude Code and Codex.

Open source

Mikie and Cline

Cline is an open-source agent you watch step by step. Mikie is a context and receipt layer for Claude Code and Codex.

Speed first

Mikie and Zed

Zed is a fast editor with an agent panel. Mikie is a context and receipt layer for Claude Code and Codex.

Adds receipts

Mikie and Claude Code

Claude Code is a capable agent on its own. Mikie hands it the exact lines a task needs and writes down what every turn cost.

Try it in shadow

Mikie and Codex

Codex already reads your repository well. Mikie narrows what it reads to what the task needs, and lets you measure the difference before switching anything on.