Insights · Measured, then shipped
We hit three provider limits in one day.
Then we went looking for the waste — and found that most of what our agents were spending, they were spending on re-reading things they already knew. This is what we measured, what we changed, and the skill we wrote out of it. The skill is free and the audit comes with it.
Version 1.0.0 · MIT licence · released 2026-08-19 · verify with
shasum -a 256 -c SHA256SUMS.txt
The numbers · with dates
What we actually measured
Every figure below was taken from our own fleet on the date shown. One of them is not finished being measured, so it is marked pending rather than claimed. We would rather publish a gap than a guess.
The waste · five sources
Where it was all going
A fixed prefix, re-read every turn
Around 35,000 tokens of tool schemas, skill descriptions and slash commands went up with every single turn. Roughly 33,000 of that was harness overhead, not our own instructions — and most agents never invoked most of those tools.
The same work, attempted again and again
Issues that could not progress were retried on a schedule. One ticket alone accumulated 75 runs without moving. Nobody had told the scheduler that some failures are not worth retrying.
Retries against a wall that was already up
Quota exhaustion was classified as a transient upstream error, so the platform retried it — against the same exhausted quota, immediately, repeatedly. The error text literally said the limit had been hit.
Failed runs, paid for in full
A run that fails after loading its full context still pays for that context. Failure is not free, and at scale it is not cheap.
Volatile fields ahead of stable ones
Prompt sections that change every turn were being assembled before sections that never change, which defeats provider prefix caching. Same content, wrong order, no cache hit.
One caveat worth stating plainly, because it changes the conclusion: the vast majority of those tokens were cache reads, billed at a fraction of fresh input. Raw token counts overstate the money. What this work really buys is headroom against provider limits and far less churn — not a large cash saving. We would rather say that than sell the bigger number.
The fixes
What we changed
- Terminal means terminal — Quota errors stop being retryable. If the provider says the limit is hit, waiting is the only strategy that works.
- Cap the re-attempts — An issue that has not moved after N runs gets parked for a human rather than looping.
- Cut the tool surface — Deny the tools an agent provably never calls. We measured which ones from real invocation logs, then proved the request actually shrank by capturing it on the wire.
- Stable sections first — Assemble the unchanging part of the prompt before the volatile part, so provider caching can do its job.
- Right-size the reasoning — Maximum reasoning effort on every routine edit is a tax with no return. Match the setting to the task.
The skill · free, MIT
CIA Optimise
One markdown file. No scripts, no hooks, no installers, no network calls. It asks five questions before any code gets written, requires a runnable check before anything is called done, and names the things that never get simplified away — validation at trust boundaries, error handling that prevents data loss, security, accessibility, and anything you explicitly asked for.
The part we have not seen elsewhere is the unattended section. Most minimalism guidance assumes a human is reading the reply. Ours runs on schedulers and boards, so the skill never re-arms itself, never closes a task it has decided is unnecessary — it escalates instead — and never shortens a required status report to save words.
Install
Save the file, then invoke it by name. Explicit invocation only — it will not fire on its own.
~/.claude/skills/cia-optimise/SKILL.mdVerify
Every file in the bundle is checksummed.
shasum -a 256 -c SHA256SUMS.txtProvenance · we had it audited
Original work, and we checked
Plenty of people have written about making coding agents build less. Two projects shaped our thinking and deserve the credit: Ponytail, which popularised the idea of an agent climbing a decision ladder before writing code, and Headroom, which compresses tool output before it reaches the model. We studied both. We did not copy either.
Because "written from scratch" is a claim, we had it tested. An independent adversarial audit was tasked with finding a copied sentence: it compared our text against both the upstream Ponytail skill and an internal build derived from it, using 8-word shingle overlap and longest-common-run analysis.
Against either baseline. The longest shared run of ordinary prose was 5 words. The audit also disclosed two structural echoes it considered worth naming rather than arguing away — both are in the report, which ships inside the download.
The CIA Code
Why it is built this way
Prove It
Every number here was measured, and each one carries its date. The one that is not measured yet says so, in the table, rather than being quietly rounded into the story.
Keep It Simple
The largest saving came from sending less, not from building more. No new dependency, no new service, no new abstraction.
Guard the Trust
None of the savings come from weakening validation, error handling, security or the audit trail. Those were ring-fenced before we started cutting.
Running a fleet of your own?
The free skill is the discipline. The Fleet Efficiency Pack is the method — the full playbook, the wire-capture technique for measuring your own stack, and the prefix-stability guide, with the numbers and how to reproduce them.