Skip to content
Back to skills

Shipping Artifacts

ASecurity

Every claim ships with the artifact that proves it — test output, a run dir, an attestation, a review verdict. Use when finishing a change or asserting that something works.

  • 3 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added September 19, 2026
ai-agentsrustgo

Security analysis

A100/100

Scanned September 19, 2026

npx -y skills add SathiaAI/adversarial-review --skill shipping-artifacts --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Shipping Artifacts?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Shipping Artifacts
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/sathiaai-shipping-artifacts/badge)](https://www.skillsdirectory.com/skills/sathiaai-shipping-artifacts)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: shipping-artifacts
description: Every claim ships with the artifact that proves it — test output, a run dir, an attestation, a review verdict. Use when finishing a change or asserting that something works.
---

# Shipping artifacts

A claim without an artifact a skeptic would accept is an opinion. This project's whole
premise is "is this change correct?" answered with *evidence*, not model self-assurance — so
hold yourself to the same bar.

## The rule
When you assert something is done, correct, safe, or passing, attach the artifact that lets
someone else confirm it **without rerunning your reasoning**:

| Claim | Artifact that ships with it |
|---|---|
| "Tests pass" | The suite's `N passed, M failed` line, and — for a change that could regress others — the byte-identical fail-set vs the base branch (prove the deltas are yours or nobody's). |
| "No regressions" | A baseline run of the *unchanged* base plus your run; a diff of the two fail-sets. |
| "The verdict is X" | The run dir under `.adversarial-review/` — the immutable audit record — and the `aggregate.py`-computed `verdict.json`, never a hand-authored verdict. |
| "It's tamper-evident" | The `compute_attestation()` digest (and a detached signature when `--sign` is used). |
| "It was reviewed" | The adversarial-review run's verdict, attached to the PR, at the stated tier. |
| "A behavior holds" | An offline test that asserts the *observable outcome and the failure path*, not a mocked success. |

## Discipline
- **Determinism.** An artifact you cite must be reproducible: same inputs → same artifact.
  No `Date.now`, unsorted iteration, or unrecorded RNG in anything recorded.
- **No secrets in artifacts.** Keys never land in run dirs, tests, logs, or docs.
- **Honest status.** If a check could not run (a missing permission, an unavailable binary),
  the artifact records **BLOCKED / unknown** — never a faked pass. A BLOCKED verdict from an
  honest gap is a correct artifact; a green one you can't back is a lie that outlives you.
- **The sandbox is not storage.** A build environment can vanish, so don't let keepable work live only
  in a scratch dir. But **push only when you have explicit authorization to push** — this repo makes that
  non-negotiable (see `AGENTS.md`). Where you may push, do it the same turn you commit; where you may not,
  keep the work safe another way (a durable clone, a patch) and hand the commit to whoever merges. The
  durable artifact is the recorded commit, not the scratch copy.

## Anti-pattern
"I ran it and it looked fine." Ship the output, the run id, the diff, or the verdict — the
thing the next person opens instead of trusting you.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.