Start

Why every step costs

Three things we learned while building the pack, in the order we learned them. Each one changed what we built.

Every step re-reads everything

A Claude Code session is a loop. Claude reads the conversation so far and takes one step: it opens a file, runs a search or makes an edit. The result joins the conversation, and then the whole conversation goes back to the model for the next step.

Caching makes that re-read cheap per word, but not free. In a session of 220,000 tokens, each step costs about four cents$0.044 before any thinking happens. So the unit of cost is the step. Making each answer a little shorter saves a little; taking out a step saves a whole re-read of everything.

Our first measurement came out exactly along that line. The words per step barely moved, and the steps fell: 5.2 steps per job with the pack, against 7.2 without5.2 against 7.2, with results as good as without it.

Watch it in the replay on the overview.

What a step leaves behind

Then we found the second half of the bill. Whatever a step reads stays in the conversation, and every later step reads it again, until the session is compacted.

In the author’s own sessions, a token added to the conversation was typically read again by 143143 later steps. Writing it costs once; over its life, re-reading it costs more than four times as much$37.

Here is one million tokens, written once and then read again by every later step:

That sank the first versions of seek: they saved steps by adding text, and the text outlived the steps it saved. The version that ships asks Claude to read less, and leaves less behind.

Claude follows orders, not advice

The natural way to change what an AI agent does is to give it advice. We tried. A hint on a search result, telling Claude which tool would answer exactly, was followed 0 times in 450 of 45. Invited to hand work to a cheaper model, Claude did it 0 times in 174 runs0 of 174.

An order in your name, arriving with your prompt, is a different thing. It becomes part of Claude’s plan before the plan exists. Here is how often Claude did what it was told, by where the instruction arrived:

Build the approved plan through a hand-offwith your approval of a plan8 of 88 of 8
Hand this feature work to the workerwith your prompt24 of 2424 of 24
Look around in few, lean stepswith your prompt44 of 4444 of 44
Hand the rest of this search offin the middle of Claude’s own search1 of 61 of 6
This tool answers your search exactlyon a tool’s result0 of 450 of 45

So the pack’s mods either answer inside what Claude already reads, like lsp-plus, or speak in your name beside your prompt, like delegate, route and seek.