Autoprompt Skill

Autoprompt Skill is a workflow you install into a coding agent rather than a program you run: one command hands it a goal together with the constraints and the definition of done, and it drives the execution loop itself — scoping, planning, building, testing, reviewing, repairing and verifying — instead of waiting to be prompted at every step. Eleven agent tools are supported, among them Claude Code, Codex, VS Code, OpenCode, Kilo Code, Prime Agent, Oh My Pi, DeepSeek Harness, Reasonix, Hermes Agent and Grok Build. Work is organised into three paths, direct, light and roadmap, chosen automatically or named before the goal, and carried out by a layered fleet that separates coordination, management, execution and independent judgment, so the author of a change is not its approver. Concurrency is capped, and model choice runs through a registry. It is MIT licensed JavaScript with 1,298 stars, 88 forks, eleven contributors and 1,235 files. The claim of 45% fewer failures rests on a single Terminal-Bench 2.1 run by the author.

Who built itThe owner account, credited with 22 of the repository’s 44 commits — 17 under the 77499101+Spielewoy address and 5 under a second Spielewoy address. Eight commits carry a co-author trailer naming Claude Opus 5, and three are attributed to an account called claude; the rest come from eight other accounts, most of them a single change, and the contributor list has eleven entries in total. The MIT licence is held as “Copyright 2026 Spielewoy”.

1. 01

   Six releases in twenty-three days, then nineteen days of quiet

   The repository was created on 2026-08-17 at 22:40 UTC and the first commit — Autoprompt Skill 1.0.0 — landed two minutes later. Six releases followed in twenty-three days: v1.0.0 on 2026-08-18, v1.0.1 and v1.0.2 both on 2026-08-19, v1.0.3 on 2026-08-20, v1.0.4 on 2026-08-21, and then a nineteen-day gap before v2.0.0 on 2026-09-09. None of them is a prerelease or a draft. The 44 commits split 29 in August and 15 in September, and the newest one, Preserve community PR authorship for the released v2 integration, is timestamped 2026-09-09 at 22:05, while the repository’s pushed timestamp is 2026-09-28, three weeks later, so some of that later activity is not visible in the commit list. Around the code sit 1,298 stars, 88 forks and 4 watchers, three open issues, an MIT licence, 1,235 files and 375,291 KB of repository.

2. 02

   The 45% number, and where its evidence stops

   The claim is on the README’s first line — a coding-agent workflow that cuts failures by 45% by reviewing, fixing and rechecking its work — with a badge reading Terminal-Bench 2.1, plus 14.61 points. OpenCode alone solved 60 of 89 tasks for 67.42%, failing 29; OpenCode with Autoprompt solved 73 of 89 for 82.02%, failing 16 — thirteen more solves and 45% fewer failures. The method page adds the setup — DeepSeek V4 Flash, OpenCode 1.18.7, the same 89 tasks — and notes these are version 1 results, with version 2 benchmarks to follow. The evidence section is where it thins. The baseline is said to have all 89 retained task verdicts at a path under benchmark/terminal-bench-2.1/; the Autoprompt side is called the recorded completed aggregate in a comparison report whose original per-task map was not retained, so that result cannot be rebuilt task by task. That benchmark/ directory is not in the published tree at all: the contents API returns 404 for it. The promised trade-off — three times the time, twice the tokens — is a planning estimate from user experience reports, because the timing and token logs were not retained. DeepSeek’s own 82.7% is called an external reference, not a comparable third run. So: no dataset in the material, no reproduction command, no surviving per-task output, no independent rerun. Read 45% as the author’s own single measurement.

3. 03

   The repository’s own benchmark gate says no

   Beside that headline sits a body of work that argues the other way. scripts/benchmark-evidence/ holds 25 files and 488 KB of harness — aggregate.cjs at 53 KB, manifest.cjs at 25 KB, route-holdout.cjs at 22 KB, a spool, a run lease, a trust registry and a quality gate — and agents/contracts/schemas/ carries twelve more benchmark schemas covering attempt evidence, mechanism evidence, run manifests, aggregate reports, pricing and result bundles, exercised by an 82 KB source test. What it does with its own results is the interesting part. The three-arm low-compute canary records one task, one repetition, evidenceClass: "harness-mechanics-only", and stamps every record it emits with qualityClaimEligible: false and comparisonClaimEligible: false; its document says flatly that this is a mechanics check, not a quality benchmark, and that its results cannot support quality, cost, latency or redesign-effect claims. The real Codex observation is carried over rather than rerun, still blocked with PROVIDER_UNSUPPORTED for codex-command-sandbox-network-open, and a real three-arm comparison is still open. The 16-row route fixture is marked synthetic, and readiness fails closed until independent human labels, two raters, inter-rater agreement, adjudication, a development and test split and a pre-tuning freeze exist; its perfect confusion matrix is mechanics evidence only.

4. 04

   Nineteen days of silence, then ten replies in one evening

   Between 2026-08-19 and 2026-08-31 the repository collected a steady stream of outside work: Grok Build support (#5), a CRLF shebang and missing exec-bit fix (#7), an Oh My Pi adapter (#8), a supervisor and model-casting addition to it (#11), two independent attempts at the non-blocking stdin crash (#15 and #24), a Python 3 resolution fix for the bash installer (#16), Hermes Agent support (#17), five grammar corrections from one contributor (#18, #19, #21, #22, #23), a report that the DeepSeek preset denied two unregistered tool names (#20), three filings of the same Windows Git Bash path failure (#12, #13, #14) and a complaint that the compatibility guide told readers to replace two bracketed tokens without saying what they stood for (#4). On 2026-08-23 the author answered several of them with the same sentence — “Currently working on some major refactors” — and then went quiet. On 2026-09-09, within minutes of the v2.0.0 release at 21:25, ten of those threads received a reply saying the issue was fixed in v2 and the reporter credited in the README, and a final commit is titled Preserve community PR authorship for the released v2 integration. The credits page names seven contributors with PR numbers and adds a disclaimer that a credit does not mean an old pull request was merged unchanged, or that a provider passed release verification.

5. 05

   Thirty-two roles, three paths, and a gate that fails closed

   The v2 rewrite replaced the 25-role fleet with a fixed set of 32 physical roles described as seven active roles plus 25 inactive compatibility aliases, projected into all eleven harnesses along with 18 framework procedures. The source README states the rule that shapes all of it: “There is no default route.” DIRECT and LIGHT need no coordinators or managers, ROADMAP may use the run coordinator and the work-group manager only when the policy admits them, the run owner keeps the goal and the final response, and the controller enforces physical launches, assignment ownership, independent checking and recovery; leaves and aliases cannot dispatch. Responsibility is layered from L0 to L4 so that an author cannot approve its own work, and the work-paths document insists that file count does not decide the route — a twenty-file rename can be direct, a three-file change across two deployed services can need roadmap. The same discipline applies to admission: “Missing required capability evidence must block admission without an unrestricted native fallback.” The bill arrives in the issue tracker: a Windows user reports every activation refused, because the bundle ships ten reviewed records that are all Linux x64 and the trusted key list is empty, against a support table that says the tested versions passed Linux runs.

6. 06

   Two shell ports and the Windows bill

   Maintaining eleven host projections is the visible cost; maintaining two operating-system ports is the quieter one. scripts/install/ exists twice over in 21 files and 947 KB — install.ps1 at 107 KB and install.sh at 74 KB, with shared libraries of 278 KB and 228 KB — and the bugs follow the seam between them. v1.0.2 was published from a Windows release runner with no .gitattributes, so the LF-committed launcher was packed with CRLF endings and no executable bit, and every fresh macOS or Linux install failed with either permission denied: autoprompt or env: node : No such file or directory; the fix adds .gitattributes and a packaging test that asserts each declared entrypoint carries an LF shebang, no carriage returns and the executable bit. A later pull request normalises Git Bash short paths through their long Windows form, which it says cleared three lifecycle failures in release-readiness run 32481402055 and unblocked the v1.0.4 publication. Reports that Windows hosts resolve bash to a WSL stub arrive three times over the same day, native Windows support is requested and refused twice more, and the still-open pull request that fixes npm .cmd shims notes that npm run verify remains blocked by six pre-existing Windows, macOS, Codex and cache-environment failures in the CLI test suite.