← All articles

Configuration

Claude Code Skills and Hooks That Pay for Themselves in Tokens

Every skill you install is a small subscription you pay in context, on every session, whether or not it fires. A hook is a different purchase altogether — and that difference is where most of the money is.

Written by Ted Mavera, Product Manager at HarnessUp, the company that makes the tool described here. The texts on this blog are written with the help of artificial intelligence and reviewed by the team before publication.

A cream ceramic machine with five skill cartridges, each dispensing orange coin discs into its own glass tube, a brass hourglass at its side and a loose cartridge in front.

Why this article exists

There are lists of Claude Code skills everywhere now: the top ten, the best for frontend, the ones that will change your workflow. What none of them tell you is that every entry on those lists is a purchase, paid in the same currency as the work itself, and charged whether or not you use it that day. The list ranks usefulness. Nobody ranks cost.

This piece does the second half. Not because skills are a bad deal — some of them are the best deal in the whole harness — but because the difference between a skill that pays for itself and one that quietly taxes every session is a difference you can measure, and almost nobody does.

Two ways to spend when you add a skill

A skill has a definition, and part of that definition is in the model's context from the moment the session starts, so the agent knows the skill exists and when to reach for it. That part is the fixed cost: paid on every session, in every conversation, independent of whether the skill ever runs. When the skill does run, its body is loaded and its instructions shape the following turns. That is the variable cost.

The fixed half is the one that gets ignored, and it compounds in the least interesting way. Ten skills that each look tiny at the bottom of a session add up to a block that is re-read on every turn of every conversation. It is the same arithmetic as any other context: what sits at the top of the conversation is charged again and again, not once.

Hooks work differently, and this is the part the lists get wrong most often. A hook is a command the harness runs on an event — before a tool call, after an edit, when the session stops. Most hooks add no tokens at all: they add latency, and they change what happens next. A hook that blocks a useless write, refuses a repeated read or compiles the project after an edit costs you milliseconds and can refund an entire turn. In token terms, the good ones are close to free; the bad ones are expensive not because they read context but because they fail, retry, or fire on everything.

How to price a skill before you keep it

Four questions, in this order. What does it add to the opening context? How many turns does it add or remove when it runs? How much output does it produce? And how often does it actually fire in a normal week?

The fourth question is the one that decides most cases. A skill that fires twice a month but rides in context every session is a subscription to something you rarely use. The same skill, if it is well written, costs a fraction of what it saves on the two occasions it fires — but only if the occasions are real.

Our panel has a screen aimed exactly at this. Optimize Harness shows the measured cost of each skill, subagent, plugin and server, with a switch to turn it off, and it deliberately refuses to guess: what was not measured gets no number. Its neighbour, Rules Auditor, does the same for the rules in CLAUDE.md, showing the fixed cost of each block that gets reloaded on every round — and it proposes a leaner file without ever editing it on its own. Both are Pro; the measurement habit they enable is worth more than any single tool in the catalog.

A small open ledger book on a workshop bench with two columns headed in pencil, a magnifying glass resting on the left page and a few brass screws scattered nearby.
Fixed cost on one page, variable cost on the other. Most lists only read the right one.

The categories that usually pay

Things that stop a re-read. The single most expensive habit in an agent session is reading the same content twice. Anything that delivers a skeleton before a body, refuses a stretch the agent has already seen, or keeps a stable summary instead of the full history is buying back context that would otherwise be re-sent dozens of times. Read Router returns the index of a large context file on the first read and the full text only if asked; Delta Read tracks which lines the agent has already seen; Hot Context slices old history into fragments that stay searchable instead of resident.

Things that cap tool output. A grep that returns six hundred matches, a test suite that prints nine hundred lines, a fetched page that arrives as forty thousand characters of navigation: all of it enters context and stays. Head Limit cuts the giant grep result per call and teaches the agent how to ask for the rest; Smart Cut delivers the first page of a large file with the total size; Output Diet and Terminal Guide keep the beginning, the end and the error lines, with the full output left on disk.

Things that prevent a wasted turn. A turn is the most expensive unit in the system, and the classic waste is the turn spent discovering something a compiler already knew. First Pass builds after the edit and pushes the error into the next turn. Conduction Write blocks the write that would leave the file identical. Way Out, when a tool fails, points to an example in the project that solved the same problem — only when there is evidence for it.

Things that shorten the answer. Response Diet cuts the warm-up text before it enters the conversation, which is the cheapest possible moment to cut it. No Comment asks the agent not to comment the obvious and measures how much new code still arrives commented, without touching comments that already exist.

The categories that usually do not

Skills that restate documentation the model already has. Persona and formatting skills, unless the format is a hard requirement. Three MCP servers that overlap, each of which inspects the project at startup. And the subtler case: a skill that is genuinely useful but is installed twice, in two versions, with two slightly different descriptions — the agent now pays for both and has to choose.

None of these is a sin. They are all rational purchases made one at a time, by someone who was not shown the receipt. The problem is not the individual skill; it is that the sum was never priced.

A 20-minute audit

List what you have installed. Open Optimize Harness and look at the measured cost per item, then switch off everything you did not use this week. Open Rules Auditor and read the fixed cost of the blocks in your context file — most people find at least one block that describes a convention the project abandoned. Run three normal sessions. Compare the opening context of the first session with the fourth.

One warning, learned the hard way: turn things off in batches of one, and keep a note of what you removed. A skill you disabled can cost more in quality than it saved in tokens, and the only way to know is to notice the week it went quiet.

What this does not promise

A lean context is not automatically a good one. Some skills earn their fixed cost by preventing three bad turns a day, and cutting them would be the worst kind of thrift. The measurement tells you the price; it does not tell you the value. That judgment stays with you, which is also why every tool here is reversible with one click and why the measurement layer is free — you should be able to look before you leap.

HarnessUp itself is not a skill, a hook, a plugin, an MCP server or a proxy. It is a local second layer that acts through the official hooks of Claude Code and Codex and never reads your token or your key. The distinction matters for this article: none of the tools above adds a definition to your opening context in the way a skill does. They work on the traffic, not on the prompt.

Price what you installed. Install HarnessUp free and see the measured cost of every skill, subagent and server in your setup.

Install free