discipline
Drop a short block at the top of a session and Claude spends less without doing less. It targets the three habits that quietly run up the bill: re-printing plans and diffs already on your screen, opening a whole file to find one line, and handing small jobs to sub-agents that each pay for a fresh context. The nearest alternative is caveman, which compresses the prose instead. Measured across 106 tasks that bought a third off the cost and gave back 7.6 points of task score, because most of its saving came from the agent investigating less rather than writing less. This one never touches your writing style, and states the boundary the other is missing: it changes how much you write, never how much you do.
/plugin install discipline@fledgeling-pluginsNeeds the marketplace added first — how to do that.
Reach for it when
Session-start operating defaults that cut token spend without cutting task success, tuned for Claude Opus 5 agentic coding. Ships a byte-stable block for the cached system prefix (Perch proxy, CLAUDE.md, or an output style) plus the evidence behind every clause.
Not for
Not a prose-compression style: compressed register is the lever this skill exists to argue against.
What ships with it
- No scripts — prose only
- 2 reference files
- Measured evals
Say any of this
- why is this burning so many tokens
Taken from the skill’s own trigger description — these are the phrases it listens for. You do not have to match them exactly.
why is this burning so many tokens
A ~220-token block for the top of a session.
It targets three habits, not your prose:
· re-printing plans and diffs already on your screen
· opening a whole file to find one line
· handing small jobs to sub-agents that each pay for
a fresh context
Measured across 106 tasks on Claude Opus 5. The nearest
alternative compresses prose instead: a third off cost,
and 7.6 points off task score — because 78% of that
"saving" was the agent doing less work.
This one changes how much you write, never how much
you do.Stop an agent buying cheap tokens with worse work. A ~220-token block for the start of a session, plus the measured reason every line of it says what it says.
The problem
The obvious way to cut an AI coding agent's bill is to make it talk less. There is a popular skill that does exactly that, and it is honest about its own numbers: about 8.5% off output tokens on real agentic work, against 65% on chat-style prose.
Run it on a real benchmark and something worse than "only 8.5%" shows up.
Across 106 tasks measured twice each on Claude Opus 5, the compressed-prose skill cut cost by a third, and cut task success by 7.6 points. Digging into where the saving came from explains why. The agent's steps per task fell by 33%. Its tokens per step fell by only 14%. So roughly 78% of the "saving" was the agent doing less work, not writing more briefly.
That is the trap. Told to spend fewer tokens, the cheapest way to comply is to investigate less, and every token metric rewards it while the work quietly gets worse.
What this does instead
It never asks for shorter prose. It targets the things that actually cost: restating what is already on screen, opening whole files to find one fact, and handing small jobs to sub-agents that each pay for a fresh context.
And it carries one clause the measurement bought:
This changes how much you write, never how much you do. Investigate, plan and verify as you otherwise would, and take the steps the task needs.
Every other line can be satisfied by simply doing less. That one says not to.
Does it actually work
Two separate checks, and they agree on the diagnosis.
The benchmark. 106 tasks, both arms, Claude Opus 5, graded by the benchmark's own rules.
| baseline | compressed prose | |
|---|---|---|
| Task score | 63.3% | 55.7% |
| Cost | $229.02 | $152.34 |
| Steps per task | 24.5 | 16.5 |
48 tasks got worse, 15 got better. On short tasks the effect vanishes (13 worse, 10 better, which is statistically nothing). On tasks taking 20 or more steps it is decisive: 34 worse, 4 better.
The blind taste test. 14 real pairs of finished work, three different AI judges from three different companies, none told which was which or what the benchmark thought.
| Judge | preferred baseline | preferred compressed |
|---|---|---|
| Claude | 11 | 2 |
| GPT | 10 | 4 |
| Grok | 8 | 6 |
| Total | 29 | 12 |
All three leaned the same way, on a coin-flip-unlikely margin.
Honest limits
- It has not been shown to save money. On a 106-task arm it scored level with no block at all (61.6% against 63.3%, p = 0.90) and cost 32.6% more. The quality half of the claim is measured and holds; the saving half is not, and that arm ran one sample per task against the baseline's two, so the cost figure is unresolved rather than settled. Read EVALS.md before switching it on expecting a lower bill.
- Whether this block beats its own previous version is not measured either.
- The persona-length research behind the size target was run on much smaller models. It is a reason to keep the block short, not proof about Opus 5.
- The compressed-prose skill this replaces is not overselling itself. Its README reports the 8.5% figure, links an independent study, and warns savings can go negative. The disagreement is with its rules, not its marketing.
Install
/plugin install discipline@fledgeling-plugins
Credit
The idea of a terse-output skill, and the honest agentic number that started this investigation, come from caveman by Julius Brussee (MIT). The independent measurement that first showed the gap between 65% and 8.5% was published by JetBrains.
Deeper
SKILL.md: why each clause earns its bytes, and what was rejectedreferences/evidence.md: every number, with what it rests onreferences/injected-block.md: the literal itselfEVALS.md: the deep half of the comparison, with the caveats stateddocs/blind-panel/: the raw verdicts and the withheld keydocs/deep-research/: the four research reports behind it