This directory holds Claude Code skills generalized from Mr Onion’s own
testing and engineering judgment, plus material Martin proposed directly
(code-structure, refactor-audit) and one ported from a sibling repo
(plan-to-specs). It’s split into two categories so the distinction stays
structural as more skills land, rather than flat and increasingly mixed —
see
docs/decisions/phase-18-skills-taxonomy-backlog.md
for why. Neither category has a flagship, and none of these are positioned
as secondary to another — see ROADMAP.md for full detail on
what each one generalizes from and how it was verified.
As of Phase 22, 12 skills are built: the 11 discussed since
docs/REIMAGINING.md first raised the question, plus
crawl-and-classify-pages, added once TFP-036’s spike (Phase 22) confirmed
introspect/crawl was real, non-duplicated capability rather than something
Playwright MCP already provided. None are pure agent judgment with nothing
to check against: each names the real case it generalizes from, and most
were run against a real target before being called done.
As of Phase 21 (TFP-051), these are real Claude Code plugin skills, not
just files to open and follow by hand. skills/.claude-plugin/plugin.json
declares both testing/ and engineering/ as skill directories, packaged
as the mr-onion-skills plugin and registered via a local marketplace
(.claude-plugin/marketplace.json) that .claude/settings.json enables
automatically for anyone with this repo trusted — no manual install step.
Invoke any skill by its plugin-scoped name, e.g.:
/mr-onion-skills:write-playwright-test
/mr-onion-skills:code-structure
Verified directly (claude plugin install, then a fresh session’s own
Skill tool — not opened and read manually): both produced their
documented worked-example output exactly, through the real invocation
path.
Skills about what a test should cover and what it actually proves. Lives in
skills/testing/.
write-playwright-test —
writes a Playwright suite for a described feature, judging which test
categories (functional, responsive, accessibility, etc.) actually apply
rather than generating one of everything. The floor skill: packages the
coverage reasoning that was already isolated in
src/lib/generate/prompt.ts.honest-coverage — after writing
or reviewing tests, states per test or per group what’s actually proven
and what remains unproven, with a Stable/Moderate/Blocked rating and real
reasoning per case — not a generic disclaimer. Generalizes the rating
scheme Emil was already doing by hand in DESKTOP_STABILITY_MATRIX.md.dont-stub-what-you-prove
— reads a test and flags when what’s mocked is the same thing the
assertion claims to verify. Generalizes TA-227, a real defect: a green
test that proved nothing because its assertion checked a stub instead of
the dependency it stood in for.find-the-hidden-platform
— audits for skipif guards, default-off flags, and monkeypatches that
make a real code path structurally unreachable by the suite, on any run.
Named explicitly as ISTQB’s structural (white-box) testing. Generalizes
TA-230, which cost real time three separate times in two days, including
a suite that stayed green while the app crashed on startup.plan-to-specs — expands a
TEST_PLAN.md into a spec tree mapped one-to-one onto case IDs, with a
--check mode that reports real coverage (“18 of 181 cases”, not a file
count) and exits non-zero when the plan and the tree drift apart. Ported
from Martin’s own generator-improvements branch on cah-pwa
(unmerged, commit 78e30de, with Claude Opus 5) rather than rebuilt from
scratch — that branch fixed a real bug that had hidden 17 implemented
tests inside files a broken parser had named .spec.js.regression-safety-net —
after a fix, separately confirms the originally-reported failure is gone
(confirmation testing) and that nothing outside the intended change broke
(regression testing). Generalizes ISTQB’s change-related testing category
directly — the one gap the ISTQB pass over this whole set turned up that
wasn’t already covered by something built or proposed.failure-mode-tests — a
sibling to write-playwright-test that writes dependency-timeout and
partial-failure tests as a matter of course, not only when asked.
Generalizes the Google SRE book, ch. 17 (Testing for Reliability).crawl-and-classify-pages
— breadth-first same-origin crawl to build a page map, then per-page
testing-semantic classification (interaction category, expected HTTP
status, requiresAuth heuristic, container labeling, embedded-resource
kind) of every actionable element and link. Generalizes
src/lib/introspect/dom-snapshot.ts and src/app/api/crawl/route.ts
(~1,200 lines) directly, after TFP-036 (Phase 22) confirmed — against
Playwright MCP’s real 73-tool list, not by assumption — that neither the
crawl nor the testing-semantic classification has any equivalent there.
The discovery step upstream of write-playwright-test on a multi-page
app it doesn’t already know.Skills about how code is structured and where it’s likely to break, as
distinct from what a test covers. Lives in
skills/engineering/.
code-structure — judges where new
code should go and whether a file is implementing rather than assembling,
before it’s written. Martin’s proposal, calibrated against
src/app/page.tsx and promoted after a real trial run — see
CALIBRATION.md.refactor-audit — audits an
existing codebase for structural drift against the same thresholds.
Also Martin’s proposal, trial-run against src/app/_hooks/
(useRun.ts: 442 lines, 17 useState calls across four unrelated
concerns) before promotion.reliability-audit —
audits external calls for missing timeouts, retries without backoff, and
unbounded queues. Generalizes the Google SRE book, ch. 21-22 (Handling
Overload, Addressing Cascading Failures).observability-check —
checks whether a feature has any way to tell, in production, that it’s
broken. Explicitly framed as honest-coverage’s sibling — same instinct,
aimed at production instead of the test run. Generalizes the Google SRE
book, ch. 6 (Monitoring Distributed Systems).Release engineering (SRE ch. 8) was deliberately not added as a twelfth
candidate — it’s already covered by the existing verified-release-pipeline
account skill, not duplicated here.
Every skill above traces to something real rather than a generic best
practice: five to Emil’s own prior incidents and write-ups (TA-227, TA-230,
DESKTOP_STABILITY_MATRIX.md), two to Martin’s own proposed drafts, one to
Martin’s own unmerged branch on a sibling repo, three to named chapters of
the Google SRE book, one directly to the ISTQB testing taxonomy, and one
(crawl-and-classify-pages) to Mr Onion’s own introspect/crawl source,
via TFP-036’s spike against Playwright MCP’s real tool list. Full
citation and verification detail for each is in
ROADMAP.md and the Phase 17-20 decision records under
docs/decisions/.
write-playwright-test isn’t auto-discovered by Claude Code’s Skill
tool from a bare top-level skills/ folder. Discovery expects
.claude/skills/ or a plugin manifest. Worth its own ticket; not fixed
as part of this roadmap.introspect/crawl (~1,300 lines of DOM exploration)… tracked as
TFP-036crawl-and-classify-pages.
Full comparison: INTROSPECT_CRAWL_SPIKE.md.Phase 19’s dogfooding (TFP-040/041/042) — running TestForge’s actual app
and write-playwright-test directly against test-assist, and comparing
both to its existing hand-written suite — was deliberately blocked on this
set being real rather than aspirational. It resumes now.