System One Skills

September 20, 2026 · View on GitHub

Give your agent the check result, with the full log one read away.

A focused skill for Devin, Claude Code, and Codex. Run a verbose test or build once, give the agent a compact result, and keep the full log locally for inspection. No model call, API key, or runtime dependency.

Skills guide · Source · SYS1

When it earns its place

  • More room for the task. Repetitive test output takes up less agent context.
  • Run the check once. Return its exit status, or report a timeout or capture failure.
  • Full evidence when needed. The private local log remains available for warnings, coverage questions, and diagnosis.

Use it for checks you already know produce long logs. Choose native tools when output is short, a useful quiet mode exists, or you need the full log anyway.1 The practical benefit is less repetitive output with an explicit result and a retrievable log. Better task success and faster completion have not been shown.

system-one-skills check --timeout-ms 900000 -- bun test

The current number

The only shipped skill, system-one-verify, presented 35.20% less validation text across 563 real outputs from the local Codex and Devin transcript corpus: 32.65% for Codex (356 outputs) and 38.90% for Devin (207 outputs). Claude Code had no qualifying validation outputs in this window, so it is shown as No result. The 28 outputs that actually crossed the compaction guard were 90.61% smaller, while 535 short outputs passed through unchanged. Every replay preservation check passed.

That is a text-size result, not a provider-token or complete-task result. The repo has no verified whole-task provider-native token-reduction percentage yet. The numeric scorecard gives the denominator, provider breakdown, preservation checks, and the reason every other candidate remains unmeasured. Earlier exploratory 82% and 3.9% figures stay in the results notes, where their narrower scopes are explicit.23

Install

Requires Node.js 20+ on macOS or Linux.

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/skills

The versioned release includes a SHA-256 checksum. Bun can install the same artifact. Use your agent’s skill directory, such as .claude/skills or .devin/skills, where appropriate. Existing modified skill files are never silently overwritten.

How it works

  1. Run once. Execute the original command and capture its output locally.
  2. Return a compact result. Keep the exit status and selected evidence, disclose omissions, and print the full log’s location.
  3. Inspect only when needed. Open the log for details the excerpt cannot answer.

The skill selects checks known from earlier runs to produce at least 8 KiB. Short output passes through unchanged; long output is reduced only when it is at least 50% and 4 KiB smaller. Do not run a check twice just to measure its size. Timeouts and capture problems are reported as failures. Required repository gates still apply. Tested preservation behavior and local processing cost are documented in the results notes.

All skills

One shipped skill. Ten research candidates. Only system-one-verify is installed. The others have No numeric result and need an executable adapter and evidence of a benefit over native tools before they join the package.

| Skill | Verified reduction | Discovery signal | What to use today | | --- | --- | --- | | system-one-verify | 35.20% less validation text across 563 real outputs; 100% preservation checks | 563 replay outputs; 35.20% is text-boundary evidence only | Available: known noisy pass/fail checks when native quiet output is inadequate | | system-one-explore | No result | 213/573 calls cross an illustrative 256-token headroom screen (37.17%); not savings | Focused native search; research must prove it preserves required locations | | system-one-ci | No result | Native run watching; pilot established no avoidable polling | Native run watching; our pilot established no avoidable polling | | system-one-diff | No result | 31/68 calls cross the same headroom screen (45.59%); not savings | Scoped native diffs; research must measure missed findings | | system-one-digest | No result | 28/224 calls cross it (12.50%); not savings | Native Git status; most observed outputs were already too small to justify another layer | | system-one-fetch | No result | 54/95 shared web calls cross it (56.84%); not an independent cohort | The agent's readable-page tool; extraction accuracy and extra savings are unproven | | system-one-research | No result | Same 54/95 shared web calls as fetch; do not add the totals | Native search and targeted reads; no separate benefit demonstrated | | system-one-triage | No result | No dedicated labeled cohort | Deterministic rules or the primary agent; no evaluated decision family yet | | system-one-writing | No result | No dedicated labeled cohort | Existing linters; inactive until there is relevant task evidence | | system-one-evolve | No result | No dedicated labeled cohort | A fixed reviewed policy; inactive until optimization can repay its cost | | system-one | No result | No dedicated labeled cohort | Direct selection; an extra router is unjustified for one available skill |

The full catalog explains the proposed benefit, native alternative, and evidence needed for each. The research log publishes positive and negative findings. New skills must save tokens on complete tasks while meeting correctness and latency requirements.

The discovery percentages in the table are not reductions. They answer one narrow question: how many observed outputs were at least 256 o200k_base text tokens, leaving a hypothetical 128-token skill overhead and 128-token margin. They assume perfect deletion, so they cannot justify installing a candidate or be compared with the 35.20% validation-text result. Fetch and research share one 95-call web proxy and must not be added together. The catalog scorecard publishes the denominators, provider split, and current decision for every entry.

Recent analysis covers a real failed check and outputs from 960 calls plus native reporter alternatives. The failure summary saved tokens but needed the full log for some details; the candidate analysis gives us reasons to keep the install small. A live Codex diagnosis comparison also counts follow-up reads and keeps failed setup attempts in the evidence record.

We are prioritizing more useful failure excerpts and focused search with needed source lines over a larger catalog. Both must beat the corresponding native workflow on total tokens, correctness and completion time before a new skill or expanded workflow earns a place in the package. The catalog's research priorities explain the specific evidence gaps.

The bounded early-error retention update in v0.4.1 stays within the existing check wrapper. Its separate source-bound artifact qualification verifies preservation and local processing cost; it neither admits another workflow nor establishes a whole-task benefit.

The whole-task benchmark harness is now the path to stronger claims: it selects relevant task episodes before outcomes, pairs each skill with the best native workflow, and reports Codex, Claude Code, and Devin separately. The current 3.9% diagnosis result remains one exploratory pair until that process produces held-out matched tasks; the harness itself has no efficacy result yet.

A new screen of 1,200 real file reads also argues against adding a generic read-reuse skill. Exact repeats accounted for only 1.1% of read-output text, before instructions or freshness checks. That is an optimistic opportunity ceiling, not a measured saving.

Commands and operating limits 
system-one-skills check -- npm test
system-one-skills check --cwd ./project --timeout-ms 900000 -- npm run build
system-one-skills check --log ./check.log -- sh -c 'npm test && npm run lint'

Arguments after -- go directly to the command. Use an explicit shell for pipes or chained commands. Keep any required host scheduler outside this command. --log reserves a new private file and refuses to overwrite an existing path.

Commands are noninteractive, with a five-minute default timeout, adjustable up to fifteen minutes, and a 64 MiB log limit. Exceeding a limit stops the command and reports a failure; incomplete capture is disclosed. Cancellation forwards the signal for cleanup, then escalates after a bounded grace period. Undrained inherited pipes produce log_incomplete=true and cleanup_uncertain=true. The in-memory suffix is limited to 256 KiB. A separate bounded capture keeps up to 2 KiB across 12 early diagnostic/context lines for failed commands, including when later output displaces them from the suffix. Gaps remain explicit; excerpts preserve the existing suffix selection and add at most 2 KiB of early evidence, for at most 6 KiB of retained source bytes. They do not promise complete diagnostic coverage. The matcher follows the full log’s observed stdout/stderr chunk order. Interleaved partial lines or characters can hide a diagnostic; it does not reconstruct separate logical streams. The early-error evidence separates current qualification from the historical v0.4 findings. Default logs use a private temporary directory. They may contain sensitive output; delete them when no longer needed.

Why “System One”?

A System One skill handles a small recurring operation so the coordinating agent can use its context for the task. Here, ordinary code processes repetitive logs. TypeSafe’s introduction to System One models and Jev is related background; this collection needs no model backend or TypeSafe account.

SYS1 is the related decision interface for applications: typed yes/no answers, choices, and scores across local models, hosted Jev, and compatible servers. System One Skills works independently; installing either project does not configure the other. SYS1 source.

Development and assessment 
bun install --frozen-lockfile
bun run check
bun bench/run-bench.ts
bun bench/assess-trials.ts private-observed-trials.json

The aggregate gate checks runtime behavior, skill footprint, immutable v0.4 historical evidence, separate source-bound current replay and runtime qualification, privacy, and package contents. The trial protocol requires a strong native baseline, unused tasks, complete token accounting, independent correctness evaluation, and measured completion time. Correctness or material latency regressions block adoption. Synthetic tests check behavior; real transcript replays measure text reduction. MIT licensed.

Footnotes

  1. Short checks can cost more. The 21 short development replays and 14 selected replays from an unused Claude/Devin cohort did not qualify. Wrapping them adds instruction overhead without reducing their output. Installing a skill can also add catalog overhead on tasks that never use it. Faster task completion and better task success remain unproven. Negative results. ↩

  2. Early text-replay result, not whole-task or billing savings. Three noisy successful logs qualified from 24 Devin development replays. Their 9,731 tokens became 931 output tokens plus 774 counted skill, discovery, and invocation tokens. (9,731 − 1,705) / 9,731 = 82.48%, rounded to 82%. Counts use o200k_base; retries, later log reads, and complete agent usage were not measured. These examples helped tune the implementation. Native quiet reporters were not compared on those historical tasks. Calculation and limits. ↩

  3. One observed diagnosis, not a qualification cohort. One historical Devin failure was diagnosed by Codex in two fresh sessions, reduced first and native second. The skill arm made four reads versus three and took 37.974 seconds versus 28.625 seconds from launcher start to exit. Cache effects were uncontrolled; the model checkpoint and ambient instructions were not fully attested. A blinded rubric scored both answers 6/6. Failed setup attempts add evaluation cost and are not subtracted from these arms. Complete results and limitations. ↩