discipline

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.

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 advertises 65% off output tokens on chat-style prose. An independent study then measured it on real agentic work and got about 8.5%.

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 13.6%. 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.

Why this instead of caveman

Both cut the agent's output. Measured on the same 106 tasks, same model, same effort:

	plain Opus 5	caveman	discipline
Output tokens	4.09M	2.41M (-41.0%)	3.43M (-16.3%)
Task score	63.3%	55.7%	61.6%
Is that score drop real?		yes, p < 0.0001	no, p = 0.90

Caveman cuts more output. It is 2.5 times better at the thing it is for. If output-token count is the only number you care about, it wins and this skill does not.

Discipline is the one that still gets the work right. It holds task score level with using no skill at all, where caveman gives up 7.6 points on a result that is not close to chance.

The reason is in where each saving comes from. Caveman's steps per task fell 32.7% while its tokens per step fell only 13.6%, so roughly 78% of what it saved came from the agent taking fewer steps: investigating less, checking less, finishing sooner. That is not compression, it is less work, and a token dashboard cannot tell the two apart. Discipline's steps fell 12%, and it carries one clause written to stop exactly that: it changes how much you write, never how much you do.

So pick on what you are optimising. If you want the biggest possible output cut and the tasks are short or you will check the work yourself, caveman is the stronger tool and its README is honest about its own numbers. If the agent is doing long unattended work you intend to trust, the 16% cut that costs nothing measurable is the better trade, and the 41% cut that costs 7.6 points is not.

Or run both, which nobody has tested. They attack different axes: caveman shrinks what gets written, this shrinks what gets re-sent, re-read and re-delegated. Running the pair is coherent, and the four-arm comparison that would show whether they compose or interfere has never been run. Until it has, "pick one" is a convenience rather than a finding.

Caveman also has one advantage this skill cannot match: it installs anywhere, across 30-plus agents, with no proxy and no assumptions about where its text lands. This one needs a cached system prefix to inject into, and without one the honest move is not to install it at all.

The score comparison is the part this measurement supports properly; see EVALS.md for where it is thinner.

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.
• The prize is small, and that is now measured rather than assumed. An independent study of 2,848 Claude Code runs puts generated output at 10.4% of the actual bill (cache reads and cache writes are about 80% of it) and finds that the share of input cost any user-side layer can reach at all is roughly 5 to 6%. So a 16% output cut is worth low single digits of total spend at the very best. This skill used to assume 14% with nothing behind it; the real number is smaller.
• A bigger study found the same trap this one did. That paper is called Token Reduction Is Not Cost Reduction, and its heaviest compression arm cut delivered tokens by 38% while the bill went up 6.8%. Its correlation between tokens saved and money saved was statistically indistinguishable from zero. That is the same shape as this skill's own result, from another lab at 27 times the scale.
• Whether this block beats its own previous version is not measured either.
• It needs somewhere cached to live. The block only pays for itself inside a cached system prefix. Without an injection point it costs full price on every turn, forever, to deliver instructions about spending less, so the right answer with no injection point is not to install it.
• 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 its mechanism. Its README carries an honest-number warning, and its own numbers file lists the aggregate output saving as unpublished. The disagreement is with its rules, not its marketing.

Install

/plugin marketplace add fledgeling-co/fledgeling-plugins
/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 rejected
• references/evidence.md: every number, with what it rests on
• references/provenance.md: the tier and source of every figure, and the gate reads it
• references/injected-block.md: the literal itself
• scripts/block-check.py: the gate. Run it; it must exit 0
• EVALS.md: the deep half of the comparison, with the caveats stated
• docs/blind-panel/: the raw verdicts and the withheld key
• docs/deep-research/: the four research reports behind it