| name | skill-artifact-testing |
| description | Methodology for testing Claude Code skills and agents ("artifacts") before they ship — a 9-layer evaluation that moves a review from "looks fine" to "provably works." Load this skill whenever someone wants to review, test, validate, or audit a skill or agent (a SKILL.md, an agent definition, or a Gerrit change that adds one), asks "is this skill/agent any good?", "does this fit the documentation review system?", "will this skill activate correctly?", or wants to verify a skill's claims against ground truth. Provides per-layer criteria, a prompt library, a one-page checklist, and a fixture template.
|
| metadata | {"version":"1.0.0"} |
Artifact Testing Methodology
This reference defines how to test a Claude Code skill or agent (an
artifact) before it ships or merges. It is the reference half of the
artifact-testing capability; the artifact-tester agent drives the workflow
and loads this skill.
Core principle
Read-throughs find typos; running it against real data finds the lies.
A skill or agent is a prompt that ships to other people and runs on real work.
Its most dangerous defect is an instruction that sounds plausible on a
read-through but is false against real input. You only catch those by
executing the artifact and verifying its claims against an independent source
— never against the artifact's own text.
Skill vs agent — what you are testing
- Skill — passive reference (
SKILL.md + optional references). Tells the
model what to know. Its description: frontmatter is its trigger: it
decides when the skill auto-loads. A bad trigger means it never fires when
needed, or fires on unrelated work.
- Agent — active worker (an agent definition). Tells the model what to
do: fetch inputs, perform steps, write files, sometimes commit. Agents load
skills as they work.
A frequent defect is category error: a multi-step workflow filed as a
passive skill (or vice versa). Layer 2 catches it.
The 9 layers
Testing is layered; each layer catches a different defect class, and the later
layers are the ones most reviews skip — which is why defects survive.
| # | Layer | Catches | Usually skipped? |
|---|
| 1 | Define the contract | Vague scope, no success criteria | Yes |
| 2 | Static review | Contradictions, miscounts, style, category error | Partly |
| 3 | System integration | Duplication, fragmentation, orphaning, drift | Almost always |
| 4 | Trigger test | Activates wrongly / never activates | Almost always |
| 5 | Dynamic run (full) | Instructions false in practice | Partly (front half only) |
| 6 | Verification & citation | Wrong results that look right | Often |
| 7 | Edge cases | Fragility, bad failure behavior | Almost always |
| 8 | Baseline comparison | Whether it actually helps | Almost always |
| 9 | Capture as fixture | Regressions on the next revision | Almost always |
A quick "is this any good?" review may stop after Layer 5. Promoting an
artifact to a documentation-review-system standard requires all nine.
Layer 1 — Define the contract first
What. Before testing, write the artifact's promise: the task, the exact
activation conditions, and what correct/complete output looks like.
Why. Without a contract written in advance, you grade intuitively and
notice only what the artifact does (confirmation bias). A contract written
first is something the artifact can fail against.
Good looks like. Three crisp statements (purpose, activation, output spec)
plus 3–5 objective pass/fail criteria.
Prompts:
Read <TARGET>. Without judging it, state its contract in three parts:
(a) the single task it performs, (b) the exact conditions under which it
should activate or be used, (c) what a correct, complete output looks like.
Then list every place the contract is ambiguous or unstated.
From <TARGET>, write 3–5 yes/no acceptance criteria a reviewer can check
objectively (e.g. "identifies 100% of new public symbols", "never flags
private/internal symbols").
Layer 2 — Static review (correctness and category)
What. Read for internal defects, then check it is the right kind of
artifact in the right shape.
Why. Two defect classes: internal (self-contradiction, miscounts, broken
cross-refs, claims unverifiable from the text) and category/shape (a workflow
filed as a skill, broken naming/frontmatter conventions).
Good looks like. Each issue quoted with its location; a clear verdict on
category and convention conformance.
Prompts:
Review <TARGET> for INTERNAL defects only: self-contradictions, miscounts,
broken cross-references, conflicting instructions, and any claim that cannot
be verified from the file itself. Quote each with its line number.
Is <TARGET> the right CATEGORY — passive reference (skill) vs active workflow
(agent)? Quote the content that proves it. Do its name, location, and
frontmatter match its siblings in <DIR>?
Layer 3 — System integration (the centralization test)
What. A skill/agent joins one centralized, coordinated documentation
review system plus a dispatcher. This layer asks: does adding it keep the
documentation review system centralized, or
fragment it?
Why. The value of the system is one place per capability, one source of
truth per rule, one dispatcher that knows what exists. Each uncoordinated
addition erodes that:
- Duplication — re-implements what a sibling already does → two sources of
truth that drift.
- Orphaning — dropped in the folder but never registered in the dispatcher,
index, or how-to → nothing knows to use it.
- Fragmentation — doesn't compose with siblings (re-derives instead of
referencing), wrong conventions, or lands in a personal/local location.
- Drift — a local copy diverges from the centralized source.
These are invisible when the file is reviewed alone. Hold it against the whole
set and the dispatcher.
Good looks like. Adds a capability no sibling provides (or supersedes one
that is then removed); references siblings rather than restating them; is
registered in dispatcher + index/README + how-to; follows shared
conventions; lives in the single canonical source with no local/central drift.
Prompts:
Treat <DIR> as the centralized documentation review system plus its dispatcher (<DISPATCHER FILE>).
For <TARGET>, report:
1. CAPABILITY OVERLAP — does any existing skill/agent already do part of what
<TARGET> does? For each, say reference / replace / merge.
2. COMPOSITION — does it re-state rules a sibling owns? Name the sibling it
should reference instead.
3. REGISTRATION — is it wired into the dispatcher routing table, index/README,
and how-to docs? List every place it is MISSING.
4. CONVENTIONS — do name, location, and frontmatter match the set?
5. VERDICT — does adding it keep the set centralized or fragment it?
Map the documentation review system in <DIR>: every skill/agent with its purpose and the
who-loads-what relationships. Place <TARGET> on the map: what it loads, what
loads it, and any duplicated capability or orphaned node.
SOURCE-OF-TRUTH drift: compare the canonical copy (<CANONICAL PATH>) against
the local install (<LOCAL PATH>) for <TARGET> and its dependencies. Report
content/version differences and what must be synced.
Layer 4 — Trigger / activation test
What. Test whether the artifact fires at the right time — and is reached by
the right caller. In the documentation review system a skill is usually
loaded by an agent (or routed to that agent by the dispatcher), not just
auto-activated by its description: frontmatter. Check both: the trigger and
the loading path.
Why. A skill that never activates is dead code; one that over-activates is
noise; one no agent loads is unreachable. The trigger is half the contract and
is almost never tested. Near-misses (same domain, out of scope) are where
triggers fail.
Good looks like. Fires on every on-target prompt, silent on every
off-target one, near-misses explainable either way.
Prompt:
Here is the description frontmatter of a skill:
--- <PASTE description only> ---
For each prompt below answer YES (should activate) or NO, using ONLY the
description. Flag any where it is ambiguous or would decide wrongly.
ON-TARGET (should fire): <3 prompts using/paraphrasing its triggers>
NEAR-MISS (judgment calls): <3 adjacent-but-out-of-scope prompts>
OFF-TARGET (should not fire): <3 unrelated prompts>
Then propose 3 prompts its wording would wrongly MISS and 3 it would wrongly
CATCH, and suggest a tightened description.
Layer 5 — Dynamic run on real input (the whole thing)
What. Run the workflow end-to-end on a real patch or source file from
its domain — including phases that write files or commit — in a scratch
copy, showing diffs instead of applying.
Why. This exposes instructions that sound correct but are false. A common
shortcut runs only the safe, early phases and skips the ones that do
something — which are exactly the ones that can cause damage.
Good looks like. Every phase executed; a phase-by-phase note of where the
instructions matched reality and where they didn't; edits shown, not applied.
Prompt:
Use <TARGET> to perform its full workflow on <INPUT>. Execute EVERY phase,
including any that write files or create commits — but make changes in a
scratch copy and SHOW diffs/commit messages instead of applying. After each
phase, state in one line whether the instructions matched what you actually
encountered. Choose <INPUT> to exercise each distinct case the artifact claims
to handle.
Layer 6 — Verification & citation
What. For every factual claim the run produces, check it against an
independent source — rendered output, the live system, the source file —
not the artifact's own description.
Why. A confident wrong answer is the most dangerous output; it passes a
glance. Verifying against reality grades the artifact and tends to surface
real bugs in the work itself. If a source cannot be reached, mark the claim
UNVERIFIED — never assume.
Good looks like. Each claim labeled CONFIRMED / REFUTED / UNVERIFIED with
the independent source quoted.
Prompt:
Take the run's output and verify each factual claim against an INDEPENDENT
source (rendered output, live page, source file) — NOT <TARGET>'s own text.
For each: CONFIRMED / REFUTED / UNVERIFIED, with the source quoted. Record
separately any real bugs in the WORK ITSELF that the run surfaced.
Notes on common verification traps:
- Numeric self-claims (lengths, counts, versions) — measure them; do not
trust the prose.
- Exact source line citations (
file.cpp:81-82) drift across refactors —
re-measure and prefer citing symbols/functions over line numbers.
- JS-rendered pages may not fetch — mark UNVERIFIED rather than assume.
Layer 7 — Edge cases
What. Run it on the inputs it would rather not see: empty, out-of-scope,
malformed, oversized, or containing only what it should ignore.
Why. "Notable" examples prove only the happy path. The question is whether
the artifact fails gracefully (clear no-op/message) or wrongly (invents
work, produces confident garbage).
Good looks like. Sensible explicit behavior on every edge input.
Prompt:
Run <TARGET> against: (a) an empty/no-op input, (b) input with ONLY items it
should ignore, (c) input missing the files/sections it expects,
(d) malformed/truncated input, (e) input just outside its stated scope.
Report each as graceful / wrong / broken. Then try to make it produce a
confidently wrong result and describe what broke it.
Layer 8 — Baseline comparison
What. Perform the same task on the same input without the artifact,
then with it, and diff.
Why. The justification for a skill is that it improves on the unaided
model. Same output with and without → ceremony. Worse output → negative value.
Value = (with) − (without).
Good looks like. Each difference attributed to the artifact, with a verdict:
adds value / adds noise / no effect.
Prompt:
Perform <TASK> on <INPUT> WITHOUT loading <TARGET>; save the output. Repeat
WITH <TARGET> loaded. Diff the two. Attribute each difference to the artifact
and label it improvement / noise / neutral. End with a one-line verdict on
whether it earns its place.
Layer 9 — Capture as a re-runnable fixture
What. Save the inputs used and the findings expected, so the next revision
can be re-tested identically.
Why. A test that lives only in a transcript can't catch a regression. A
fixture turns "tested once" into "re-testable on every change."
Prompt:
Produce a fixture from this session: the exact inputs (with stable references),
the expected findings for each, the acceptance criteria from Layer 1, and a
short re-run instruction. Anchor re-runs by claim text, not line numbers, so
they survive edits.
One-page checklist
CONTRACT
[ ] Purpose, activation, output spec written BEFORE testing
[ ] 3–5 objective pass/fail criteria defined
STATIC
[ ] No contradictions, miscounts, broken refs, unverifiable claims
[ ] Correct category (skill vs agent); conventions match siblings
SYSTEM INTEGRATION (centralization)
[ ] No capability overlap with a sibling (or it replaces one)
[ ] Composes: references siblings instead of restating what they own
[ ] Registered in dispatcher + index/README + how-to
[ ] Single canonical source; no local-vs-central drift
TRIGGER
[ ] Fires on all on-target prompts; silent on all off-target
[ ] Near-miss prompts explained either way
DYNAMIC
[ ] Full workflow run on real input — EVERY phase, incl. write/commit (sandboxed)
[ ] Each distinct case the artifact claims to handle was exercised
VERIFICATION
[ ] Every output claim CONFIRMED / REFUTED / UNVERIFIED vs an independent source
[ ] Numeric claims measured; line-number citations re-checked
EDGE CASES
[ ] Empty / ignore-only / missing / malformed / out-of-scope tried
[ ] Failure behavior graceful, not confidently wrong
BASELINE
[ ] Same task run WITHOUT vs WITH; verdict: value / noise / neutral
CAPTURE
[ ] Inputs + expected findings saved as a re-runnable fixture
Fixture template
# Fixture: <artifact name>
Artifact under test: <path>
Tested revision: <commit / version>
Date: <YYYY-MM-DD>
## Inputs
| ID | Stable reference | Why chosen (branch/case it exercises) |
|----|------------------|----------------------------------------|
## Expected findings
| Input | Expected outcome |
|-------|------------------|
## Acceptance criteria (Layer 1)
- [ ] <objective pass/fail>
## Re-run instruction
<one paragraph; anchor by claim text, not line numbers>
## Baseline note
With-vs-without verdict: <value / noise / neutral>
Worked examples (defect classes this finds)
- Wrong illustrative content — a column-counting skill whose own examples
asserted "exactly 80 characters" on lines that measured 75/90; found by
Layer 6 numeric measurement, invisible to a read-through.
- Stale coordinates + triplicated rule — a skill citing
config.cpp:81-82
(actually 83–84) and docparser.cpp:1784-1788 (actually ~1813–1830), with
its alt-text wrapping rule duplicated across three skills; found by Layer 6
and Layer 3.
- Category error + fragmentation — a multi-phase workflow filed as a
passive skill, duplicating the dispatcher's fetch protocol, restating rules
other skills own, and unwired from the dispatcher; found by Layers 2 and 3.
In every case the headline defect came from Layers 6 and 3 — verification
against ground truth, and system integration — the two layers most reviews skip.
