All observed providers
35.20%1,163,594 → 754,006 UTF-8 bytes across 563 outputs.
35.20% less validation text.
Keep the full log.
System One Skills gives coding agents a compact result from a known noisy check, with the full log available for follow-up. The headline is a measured text-size result, not a billed-token guarantee.
Every candidate is accounted for. Four discovery screens show where enough output exists to test; none is presented as a saving until a paired native comparison proves correctness.
One shipped skill. Ten research candidates. Start with your tool’s quiet mode when it preserves the evidence you need. Add the wrapper when retained logs and explicit capture limits help.
One skill for known noisy checks.
system-one-verify is shipped in version 0.4.1. Use it when a check is already known to produce at least 8 KiB of output and you need its pass/fail result. No model call or API key is involved.
- Before
- The agent receives a long result full of passing-test lines or repeated build output.
- With the skill
- The same command runs once. The agent gets the exit status, a bounded excerpt, an explicit omitted-byte count, and the saved log path.
- When detail matters
- Read the relevant lines from the private local log. Follow-up reads cost tokens and can erase the initial saving.
Use it for a status decision.
A known noisy test or build has finished; the agent needs to know whether the required check passed and keep the evidence available.
Use native tools for diagnosis.
Prefer a useful quiet reporter for routine checks, and focused log reads when investigating failures, warnings, or coverage. Do not run a check twice just to choose.
The tested contract: single execution, exit status, complete captured log, exact short-output passthrough, and explicit timeout or capture failure. Twenty-five runtime tests exercise these behaviors. They do not prove better diagnoses or faster tasks. Runtime evidence.
35.20% less presented validation text.
The current scorecard replays 563 real validation outputs at the tool-result boundary. It reports exactly what was measured, who had qualifying evidence, and where the claim stops.
Compacted outputs
90.61%28 outputs crossed the guard; 535 short outputs passed through unchanged.
Preservation
0 failuresEvery replay preservation invariant passed.
Whole-task provider tokens
No resultNative-token reduction is not established for this skill.
Scope: these are UTF-8 text bytes presented at the tool-result boundary. They are not provider-native tokens, billed cost, or complete agent-task usage. A fresh paired native-versus-skill study with provider counters, correctness, and latency is still required.
Public source: scorecard methods and provider breakdown · immutable aggregate report.
Scroll sideways to compare all providers.
| Provider | Outputs | Text before → after | Reduction | Preservation failures |
|---|---|---|---|---|
| All observed providersAggregate scorecard | 563 | 1,163,594 → 754,006 bytes | 35.20% | 0 |
| CodexQualifying validation outputs | 356 | 688,245 → 463,552 bytes | 32.65% | 0 |
| DevinQualifying validation outputs | 207 | 475,349 → 290,454 bytes | 38.90% | 0 |
| Claude CodeNo qualifying output in this window | 0 | — | No result | — |
Every other skill: demand signal, not a saving
These numbers count real tool outputs that were at least 256 o200k_base text tokens. The screen imagines a 128-token skill cost, a 128-token margin, and perfect deletion. It is a way to choose the next experiment; it is not a measured reduction, task result, or correctness claim.
| Candidate | Observed calls | Calls with 256-token headroom | Decision |
|---|---|---|---|
system-one-explore | 573 | 213 (37.17%) | Test focused search and location recall |
system-one-ci | 128 | — | Native run watching preferred |
system-one-diff | 68 | 31 (45.59%) | Test hunk coverage and defect recall |
system-one-digest | 224 | 28 (12.50%) | Start with native Git porcelain |
system-one-fetch | 95 shared web calls | 54 (56.84%) | Shared proxy; test extraction fidelity |
system-one-research | 95 shared web calls | 54 (56.84%) | Same cohort as fetch; no separate result |
system-one-triage, system-one-writing, system-one-evolve, system-one | — | — | No dedicated labeled cohort |
All ten remain research-only with No numeric result. Fetch and research share one denominator and must not be added together. The full scorecard includes provider splits, native controls, missing-output accounting, and the qualification rules.
Try native quiet output first. A useful quiet reporter can beat an extra wrapper. The skill earns its place only for known noisy checks where the retained result and full log repay the local overhead.
Historical exploratory figures: 82% and 3.94%
Earlier development work reported an 82.48% smaller initial result across three successful historical Devin logs, including counted skill overhead. A separate completed Codex diagnosis pair recorded 3.94% fewer input-plus-output tokens but took 9.3 seconds longer. These are historical exploratory figures with narrower denominators; they do not change the current scorecard and do not establish billed-token savings.
The 82% calculation was (9,731 − 931 − 774) / 9,731 and only three of 24 development replays met the 8 KiB selection rule. The Codex pair had uncontrolled order and cache effects, and follow-up reads can erase an initial saving. Historical results notes.
Where the wrapper adds cost
Short output passes through unchanged. Always-visible discovery and follow-up reads can add context use when the skill is never invoked or when diagnosis needs the full log. Across the measured synthetic wrapper pairs, local processing added overhead; that is not a speedup.
The scorecard is a replay of a private one-developer convenience corpus. Only aggregate counts and a digest are public; prompts, paths, outputs, and session IDs remain private. Benchmark protocol.
Ten ideas still have to earn their place.
None of these names is installed or available as a command. Each has a proposed payoff and a native alternative to beat. Some are deferred because another layer is not justified yet.
system-one-explore
Research
Proposed payoff
Find the right repository locations with fewer reads.
Most observed searches were small. Needs a task where fewer reads preserve every required location.
Use today
Focused rg queries, path filters, then relevant lines.
system-one-ci
Research · use native watching
Proposed payoff
Avoid extra agent turns while one CI run is unchanged.
The identity pilot did not establish avoidable polling. Another watcher has no demonstrated benefit.
Use today
gh run watch for the exact run and commit.
system-one-diff
Research
Proposed payoff
Make large reviews navigable without losing relevant changes.
Smaller patches can hide defects. Must measure missed findings and all follow-up reads.
Use today
git diff --stat, then complete relevant hunks.
system-one-digest
Deferred
Proposed payoff
Combine repeatedly needed repository-state fields into one result.
Routine status is already compact. No useful repeated multi-query task has been established.
Use today
git status --porcelain and only the requested log or diff.
system-one-fetch
Deferred
Proposed payoff
Remove page boilerplate while retaining evidence and provenance.
Must preserve tables, code and citations, and beat the maintained native reader.
Use today
The agent’s readable-page tool with targeted find/open.
system-one-research
Deferred
Proposed payoff
Collect useful excerpts from named sources with fewer repeat reads.
No complete-task advantage is measured. Extraction must preserve contradictions and citation context.
Use today
Native search and source reads, tracked in a simple evidence table.
system-one-triage
Deferred
Proposed payoff
Route a narrow decision to a cheaper process, with an abstain path.
No labeled decision family is qualified. Classifier cost and mistakes can outweigh savings.
Use today
Deterministic repository rules; the main agent for judgment.
system-one-writing
Inactive
Proposed payoff
Locate mechanical draft issues before semantic editing.
No suitable writing-task cohort. Extra critique is not justified without useful findings.
Use today
An existing text linter or focused search for the named issue.
system-one-evolve
Internal research only
Proposed payoff
Improve a recurring routing policy enough to repay its evaluation cost.
First needs a useful fixed policy and untouched holdout. Search overhead may never pay back.
Use today
A fixed, reviewed rule and static evaluation set.
system-one
Inactive
Proposed payoff
Choose the smallest useful skill, including no skill.
One shipped skill does not justify another routing layer.
Use today
Expose the one relevant skill and choose native tools directly.
The next useful experiment: keep every reported failure location and a full-log line index in a compact failure manifest. The current summary omitted six of seven locations; the separate diagnosis comparison needed an extra read, though it did not show that the omissions caused it. Focused search with needed source lines is the other priority. Both still have to beat native tools on a correct completed task.
No candidate has demonstrated better complete-task token use, reliability, or speed. Admission needs a defined task, independent correctness, an adequate native baseline, all overhead and follow-up costs, and measured completion time. Full contracts and experiments · Admission protocol.
Benchmark protocol: results pending. We select relevant task episodes first, compare the best native workflow with a frozen skill, and report Codex, Claude Code, and Devin separately. Read the benchmark protocol.
What the file-read screen found
A retrospective screen of 1,200 file reads found 74 reads repeating the exact query and output within the same session. Those repeats contained 12,485 of 1,143,648 observed read-output text tokens: just 1.09% before any reuse overhead or freshness checks. This is an optimistic deletion ceiling, not measured savings. Necessary rereads cannot be assumed redundant, and no read-cache skill is admitted. Screen methods and limits.
Install the one shipped skill.
For Devin, Claude Code, and Codex. Version 0.4.1 installs only system-one-verify. Node.js 20+ · macOS or Linux.
Install System One Skills
npm install --global https://github.com/hraness/system-one-skills/releases/download/v0.4.1/system-one-skills-0.4.1.tgz
system-one-skills install-skills --target .agents/skillsUse .claude/skills or .devin/skills for those agents. Keep your repository’s scheduler and required checks.
Run a check already known to be noisy
system-one-skills check --timeout-ms 900000 -- bun testSys1 provides the separate typed-decision interface for local models and hosted services. System One Skills runs deterministic workflows without a Sys1 installation. Installing either does not configure the other.
Agent-authored analysis · Source checked September 20, 2026 at 6f69658. The scorecard is scoped to presented UTF-8 text; candidate headroom screens are not savings claims; no universal token, bill, reliability, or speed claim is made.