skills/ is a growing set, not a single product with supporting material.
write-playwright-test is the
first entry because it was the one already isolated in the codebase — it is
not a flagship, and nothing below is positioned as secondary to it. Each
entry below is a candidate: recorded so the direction doesn’t quietly
collapse into “one skill is the product” by default, not a commitment to
build it on any particular timeline. See
docs/decisions/phase-17-skills-extraction-backlog.md
and
docs/decisions/phase-18-skills-taxonomy-backlog.md
for the backlogs these came from, and docs/REIMAGINING.md for the source
proposal.
skills/ is split into two categories, skills/testing/ and
skills/engineering/ — see README.md for what distinguishes
them.
Skills about what a test should cover and what it actually proves. Lives in
skills/testing/.
honest-coverageAfter writing tests, state plainly what they do not prove — not just that they pass.
Generalizes: Emil’s own prior writing in DESKTOP_STABILITY_MATRIX.md and
cah-pwa/tests/README.md, both of which already do this by hand for a
single project. The skill would produce a stability-matrix-shaped output —
every case rated, every limitation named — rather than a one-off document.
Built as skills/testing/honest-coverage/SKILL.md
by TFP-043, porting the Stable/Moderate/Blocked scheme and real excerpted
rows from DESKTOP_STABILITY_MATRIX.md plus cah-pwa/tests/README.md’s
“a reviewer who counts spec files” framing.
Status: built.
dont-stub-what-you-proveA review skill. Reads a test and checks whether the thing being asserted is the same thing being mocked.
Generalizes: TA-227, a defect Emil found and named in his own testing work — a green test that proves nothing, because the assertion checks a stub instead of the real dependency.
Built as skills/testing/dont-stub-what-you-prove/SKILL.md
by TFP-044, porting TA-227’s _start_capture example verbatim from
test-assist/docs/ta227-228-e2e-testing-brief.md.
Status: built.
find-the-hidden-platformAudits a repo for code paths that are structurally unreachable by its own
test suite — skipif guards, default-off flags, monkeypatches that replace
the thing under test — and reports them.
Generalizes: TA-230, which cost Emil time three times in two days, including a suite that stayed green while the app died on startup.
Built as skills/testing/find-the-hidden-platform/SKILL.md
by TFP-045, from a full extraction of TA-230 into
testing/find-the-hidden-platform/TA-230.md (moved there from docs/ by TFP-255). States explicitly, in its own text,
that this is ISTQB’s structural (white-box) testing category — code
coverage of unreached paths — rather than leaving the connection implicit.
Status: built.
plan-to-specsExpands a TEST_PLAN.md into a spec tree mapped one-to-one onto case IDs,
with two-way drift detection via --check — reports real coverage (e.g.
“22 of 181 cases”) and exits non-zero on drift between the plan and the
tree.
Generalizes: an existing tool, already built twice (cah-pwa/scripts/ and
the generator-improvements branch). This is generalizing what exists, not
designing something new.
Built as skills/testing/plan-to-specs/SKILL.md
by TFP-046, porting generator-improvements (cah-pwa, commit 78e30de,
Martin Hugo + Claude Opus 5) rather than rebuilding from a description —
generalized past that branch’s hardcoded cah-pwa paths and case-ID scheme
so it runs against any repo’s TEST_PLAN.md. Verified against cah-pwa’s
own real plan and spec tree: reports real coverage (“18 of 181”), and
correctly catches a deleted spec file on --check, exiting non-zero.
Status: built.
regression-safety-netAfter a fix, confirm the originally-reported failure is actually gone (confirmation testing); separately, confirm the change didn’t break anything outside what it was meant to touch (regression testing).
Generalizes: ISTQB’s change-related testing. Unlike the four candidates above, not generalized from any of Emil’s own prior work — a fresh candidate, so there’s no cross-repo material to port.
Built as skills/testing/regression-safety-net/SKILL.md
by TFP-047, keeping confirmation and regression as two separately-testable
checks rather than one blended verdict, per its own acceptance criteria.
Status: built.
Skills about how code is structured and where it’s likely to break, as
distinct from what a test covers. Lives in skills/engineering/, holding
code-structure and refactor-audit (below) since TFP-039.
code-structure and refactor-auditDeciding where new code goes before it’s written, and auditing an existing
codebase for structural drift. Promoted into
skills/engineering/ by TFP-039, after code-structure’s
thresholds were calibrated against src/app/page.tsx and refactor-audit
was trial-run against src/app/_hooks/ (printed, not filed). Full output:
skills/engineering/CALIBRATION.md.
Generalizes: cah-pwa/skills/’s own template, applied to a useRun.ts
review (442 lines, 17 useState calls in four unrelated groups).
Status: shipped.
reliability-auditAudits external calls for missing timeouts, retries without backoff, and unbounded queues — “what happens when this dependency is slow or down.”
Generalizes: the Google SRE book, ch. 21-22 (Handling Overload, Addressing Cascading Failures).
Built as skills/engineering/reliability-audit/SKILL.md
by TFP-048, following refactor-audit’s own issue-template convention
(docs/WORKING_AGREEMENTS.md’s brief shape) for consistency, with a
three-signal table (missing timeout, retry without backoff, unbounded queue)
in place of refactor-audit’s structural-drift signals.
Status: built.
failure-mode-testsA sibling to write-playwright-test that covers dependency-timeout and
partial-failure scenarios as a matter of course, not only when the user
happens to mention them.
Generalizes: the Google SRE book, ch. 17 (Testing for Reliability).
Built as skills/testing/failure-mode-tests/SKILL.md
by TFP-049, verified against write-playwright-test’s own login-page worked
example: proposes timeout, connection-abort, and malformed-response tests
that output doesn’t cover.
Status: built.
observability-checkChecks whether a feature under test has any way to tell, in production,
that it’s broken — extends honest-coverage’s “what a green run doesn’t
prove” into “and would anyone notice.”
Generalizes: the Google SRE book, ch. 6 (Monitoring Distributed Systems).
Built as skills/engineering/observability-check/SKILL.md
by TFP-050, cross-referencing honest-coverage directly as its sibling —
same instinct, aimed at production rather than the test run — rather than
sharing the theme by coincidence.
Status: built.
Release engineering (SRE ch. 8) is deliberately not proposed as a new
candidate here — it’s already covered by the existing
verified-release-pipeline account skill.
crawl-and-classify-pagesBreadth-first same-origin crawl to build a page map, then per-page
testing-semantic classification of every actionable element and link —
interaction category, expected HTTP status, a requiresAuth heuristic,
container/section labeling, embedded-resource kind, link-type. The
discovery step upstream of write-playwright-test on a multi-page app it
doesn’t already know the shape of.
Generalizes: src/lib/introspect/dom-snapshot.ts and
src/app/api/crawl/route.ts directly (~1,200 lines), reproducing their
exact algorithms and heuristics (the BFS queue/dedup rules, the role
table, the container-label ancestor search, the interaction-category and
expected-status regex chains, the link-type and requiresAuth checks)
rather than a generic “explore the page” instruction.
Built as skills/testing/crawl-and-classify-pages/SKILL.md,
directly off TFP-036’s (Phase 22) finding that neither half of this logic
has any equivalent in Playwright MCP’s real 73-tool list — see the
Known-gaps entry below and INTROSPECT_CRAWL_SPIKE.md
for the full comparison this skill is built from. Not yet run against a
live target with a connected browser-automation tool — same method caveat
as the spike itself, stated in the skill rather than hidden.
Status: built, unverified live (see caveat above).
skills/.claude-plugin/plugin.json)
declares both skills/testing/ and skills/engineering/ via the
skills field’s array support, registered through a local marketplace
(.claude-plugin/marketplace.json) and a committed
.claude/settings.json (extraKnownMarketplaces + enabledPlugins) so
every trusted collaborator gets it, no manual install step. Not a
restructure of skills/’s own testing/engineering split — the canonical
files stayed where Phase 18 put them (two flat files,
code-structure.md/refactor-audit.md, did need moving into
<name>/SKILL.md form with real frontmatter, since that exact shape is
what Claude Code’s plugin scanner requires — a fix the ticket didn’t
originally anticipate, found by reading the reference doc rather than
assuming the existing layout already qualified). See
skills/README.md for the actual invocation syntax.introspect/crawl (~1,200-1,300 lines of DOM exploration) does not
duplicate Playwright MCP — spiked and now packaged as
crawl-and-classify-pages,
not still an open placeholder.
TFP-036 (Phase 22) ran the comparison docs/REIMAGINING.md asked for:
crawl/route.ts’s BFS same-origin crawl has no equivalent anywhere in
Playwright MCP’s real, published 73-tool list — browser_navigate is
explicitly single-page, and no tool follows links automatically.
dom-snapshot.ts’s element-role/name discovery has a real, arguably
better partial equivalent in browser_snapshot’s accessibility-tree
read, but its actual value — testing-semantic classification (interaction
category, expected HTTP status, requiresAuth, embedded-resource kind,
link-type, container labeling) — has no equivalent at all, since a
general-purpose browser-automation surface has no reason to make those
judgments. Not covered, on both halves, not a close call. Method
caveat, stated plainly: no live browser-automation tool was available in
the session that ran this spike, so the comparison is built from a full
read of Mr Onion’s own code plus Playwright MCP’s real tool
documentation (fetched directly, all 73 tools), not a live run against
the two originally-intended target pages. Full comparison, including the
per-tool breakdown: skills/INTROSPECT_CRAWL_SPIKE.md.
The remaining open piece is the live-verification gap the spike itself
named: the skill hasn’t been run yet against a real connected
browser-automation tool.