test-value: 9,967 Bytes of Test-Maintenance Judgment, With Little Outside Evidence
The standalone repo for the test-value skill has 4 stars. The project it was carved out of, tt-a1i/archify, has 87,620. That gap is the story: real provenance, because it was written to tame one codebase's test suite, and almost no public track record outside that codebase. The SKILL.md is 9,967 bytes of plain Markdown that helps an agent decide which tests to add, reuse, rewrite or delete.
Why this skill and not the repo's famous one: Archify's flagship skill ships large Node scripts, and we don't review skills that carry executable code. test-value is Markdown plus a 320-byte YAML file. I read the vendored copy end to end and diffed it against the standalone repo; a scan for network, shell and credential access came back empty. Nothing was run.
TL;DR / Key Takeaways
- The test-value skill's SKILL.md is 9,967 bytes at commit 914255b (2026-10-16), about 2,492 estimated tokens at 4 bytes per token. With its two reference files the loaded set is 23,683 bytes, about 5,921 estimated tokens. This is file-size arithmetic, not a measured session.
- test-value is MIT licensed, written by tt-a1i, and its standalone repository had 4 stars, 0 forks and 0 open issues on 2026-10-23, two weeks after it was created on 2026-10-09.
- It was extracted from tt-a1i/archify (87,620 stars, MIT, 5 SKILL.md files). Inside that repo a vendored copy is marked
metadata.internal: true, so the Skills CLI hides it from the default install list. - GitHub search finds 178 pull requests in tt-a1i/archify mentioning "test-value" (88 merged). The repo's PR template and CONTRIBUTING.md prompt for it, so that count shows a project rule, not organic adoption.
- The skill directory is pure Markdown plus a 320-byte YAML file, with no scripts. I found no published benchmark showing it improves a test suite and no Hacker News thread about it.
What the skill actually says
The opening paragraph carries the whole thesis:
"Test counts, coverage percentages and green runs are evidence, not substitutes for product judgment. When an AI writes both implementation and tests, check whether they share the same mistaken assumption."
Every test the agent adds or changes has to answer three questions: which concrete mistake would fail it, whether an equivalent refactor would also fail it, and which existing check already catches the same mistake. An eight-row table maps evidence to a decision: rewrite tests that freeze automatic output, consolidate duplicates guarding the same failure, keep a costly test that protects a distinct risk.
The part worth copying even if you never install it is the section on whether a test can tell good from bad. Relaxing an assertion? A valid alternative should pass and the original bad result should fail. Fixing a bug? Reproduce the failure first. And one line to tape to a monitor:
"A successful command with no relevant tests is not evidence that the behavior passed."
It also tells the agent to read your real test commands and CI config and pick the smallest set of checks that covers the risk. Destructive probes run only "in temporary copies or controlled fixtures." No network calls, nothing touching credentials.
What it costs you in context
The arithmetic is bytes divided by 4, a rough rule of thumb; real tokenizers differ and we measured no sessions. The 288-byte description (about 72 tokens) is the standing cost if your host lists skills. The body loads when the skill triggers, and the two references are conditional, so a small bug-fix session should land near the SKILL.md figure and a full suite audit near the larger one.
| File | Vendored in archify (bytes / est. tokens) | Standalone tt-a1i/test-value (bytes / est. tokens) | Loaded when |
|---|---|---|---|
SKILL.md | 9,584 / ~2,396 | 9,967 / ~2,492 | Skill triggers |
references/audit.md | 3,710 / ~928 | 5,299 / ~1,325 | Auditing existing tests |
references/decision-examples.md | 5,169 / ~1,292 | 8,417 / ~2,104 | Specific situations listed in SKILL.md |
agents/openai.yaml | 320 / n/a | 320 / n/a | Codex UI metadata, not instructions |
LICENSE | 1,063 / n/a | 1,063 / n/a | Never |
| SKILL.md + both references | 18,463 / ~4,616 | 23,683 / ~5,921 | Worst case, full audit |
| Whole skill directory | 19,846 bytes | 25,066 bytes | n/a |
What the community reports
Short version: very little, nearly all of it from the author's own repos. Hacker News has no thread about test-value; its only "Archify" story, item 49594377 (2026-09-07), has 1 point and one comment and concerns the parent project. The three Archify issues matching the term don't discuss the skill.
| Source | Date | What was reported | Link |
|---|---|---|---|
| GitHub repo stats, tt-a1i/test-value | Checked 2026-10-23 | 4 stars, 0 forks, 0 open issues, 4 commits since creation on 2026-10-09; last push 2026-10-16. | tt-a1i/test-value |
| PR #799 in tt-a1i/archify (maintainer) | Merged 2026-10-09 | Maintainer vendored the skill (+659/-0, 17 files) with metadata.internal: true so public installs expose only archify, and states 13 vendored files match the pinned source apart from that. A maintainer claim; I didn't diff against the pinned commit. | PR #799 |
| GitHub PR search for "test-value" in tt-a1i/archify | Checked 2026-10-23 | 178 PRs mention it (88 merged, 67 open, 23 closed unmerged) from 48 authors. The PR template and CONTRIBUTING.md prompt for it, so this is compliance with a repo rule, not word of mouth. | PR search |
| PR #848, contributor rudycelekli | Opened 2026-10-10 | Says it applied test-value and added one test; temporarily removing the fitting floor leaves all 86 existing geometry tests green while the new test fails. A concrete use of the discrimination check. Still open when I looked. | PR #848 |
Dissent is absent, and absence isn't approval. PR #848 is the most useful signal, since it shows an outside contributor following the skill's own method, but one contributor in a repo that asks for it is weak evidence. The 87,620 stars belong to Archify the diagram tool, not this skill.
How to install it
Commands are from the skill repo's README at commit 914255b. I read them; I didn't run them.
# list what's in the repo without installing
npx skills add tt-a1i/test-value --list
# install the skill
npx skills add tt-a1i/test-value --skill test-valueFor a manual install the README says to place SKILL.md and references/ in your host's skill directory. Codex users invoke it with $test-value.
Don't run npx skills add tt-a1i/archify expecting to find it. The vendored copy has metadata.internal: true, and the Skills CLI source says internal skills are hidden unless INSTALL_INTERNAL_SKILLS=1 is set. I didn't test that behavior.
Where it earns its keep, and where it doesn't
It earns its keep on a suite that's already hurting: slow CI, snapshots that go red on every valid refactor, near-duplicate fixtures, or tests written by the same agent that wrote the code.
It doesn't earn much on a small greenfield project, or for anyone who wants a mechanical checklist. Much of the text runs on judgment words like "proportionate" and "where practical," which an agent can nod at and move on. The description also lists "choosing tests during implementation," which could match many ordinary turns, though it says to keep routine use scoped to the affected behavior. That's my reading, not a reported problem.
The criticisms
- No evidence behind it. The skill tells the agent: "Claim speedups only from measurements." Its own repo publishes none: no before/after figures in the README, no eval or benchmark directory (11 files, 44,694 bytes in total). The README's story of refactors turning dozens of tests red is plausible, but it's the author's account and there's no second source.
- The examples are one project's problems.
decision-examples.mdleans on a diagram renderer: automatic route coordinates, "22 icons are each rendered through five entrypoints," and (standalone repo only) a Windows 8.3 path alias. The principles transfer; a team with a CRUD app will have to translate the cases. - The self-editing clause invites drift. A "Learn from use" section says "you may refine this skill in place." That's 1,905 of the vendored copy's 9,584 bytes, about 20% of the file. The guard rails are sensible (evidence required, tell the user, no publishing to other installs), but installs can diverge, and the vendored copy already has: a week after vendoring it was two commits and 5,220 bytes behind the standalone repo.
FAQ
What does the test-value skill do?
It helps an agent decide which tests to add, reuse, rewrite or retire, by asking which concrete mistake each test would catch and whether a valid refactor would break it. SKILL.md is 9,967 bytes at commit 914255b, MIT licensed.
How much context does the test-value skill use?
By file-size arithmetic only (bytes divided by 4, not measured): about 2,492 tokens for SKILL.md, about 5,921 with both references (23,683 bytes). The references load conditionally.
Why doesn't test-value appear when I install tt-a1i/archify?
The vendored copy carries metadata.internal: true, so the Skills CLI hides it unless INSTALL_INTERNAL_SKILLS=1 is set. Install from the standalone repo: npx skills add tt-a1i/test-value --skill test-value.
Is there independent evidence that test-value improves a test suite?
Not that I could find. The standalone repo had 4 stars, 0 open issues and no benchmark directory on 2026-10-23. The 178 Archify PRs mentioning it reflect a PR-template rule, not independent adoption.
Skill by tt-a1i, MIT licensed (Copyright (c) 2026 tt-a1i). Source: github.com/tt-a1i/test-value; vendored copy: tt-a1i/archify, .agents/skills/test-value. Quoted lines are short excerpts for review purposes. Reviewed October 23, 2026 from source files at commit 914255b of the standalone repo; no skill session was run and no costs were measured. Community reception shifts over time, and every figure above reflects what was visible on that date.