diff --git a/.agents/skills/classify-cpan-failures/SKILL.md b/.agents/skills/classify-cpan-failures/SKILL.md
new file mode 100644
index 000000000..1a7534528
--- /dev/null
+++ b/.agents/skills/classify-cpan-failures/SKILL.md
@@ -0,0 +1,69 @@
+---
+name: classify-cpan-failures
+description: Classify PerlOnJava CPAN tester failures and decide whether they warrant a GitHub issue. Use for FAIL or REGRESS records from cpan_random_tester.pl, not for implementing an already-confirmed compiler bug.
+---
+
+# Classifying CPAN Tester Failures
+
+Use the archived run log and its `FAILURES.tsv` or `REGRESSIONS.tsv` row as
+the source of truth. Verify that the reported module is the module whose test
+block actually failed; nested CPAN dependency output can otherwise make an
+association ambiguous.
+
+## Issue threshold
+
+Create or recommend a PerlOnJava issue only when all of these hold:
+
+1. The relevant upstream test suite can pass completely under system Perl in
+   the same usable environment.
+2. The PerlOnJava failure is reproducible and attributable to PerlOnJava,
+   rather than a tester timeout, missing service, missing display, platform
+   prerequisite, or unsupported native dependency.
+3. For a distribution containing XS/native code, it is an important ecosystem
+   dependency: establish this with concrete reverse-dependency evidence (for
+   example, a widely used foundational module such as Moose). Do not open an
+   issue merely because an XS distribution fails.
+
+If system Perl cannot pass all tests, do not create an issue. Record the
+environmental or upstream reason instead.
+
+Exception: when the failure is independently attributable to a PerlOnJava
+internal bug, create or recommend an issue even if the full upstream suite
+cannot run cleanly. First reduce it to a small, self-contained reproducer
+that can become a project regression test when the bug is fixed.
+
+A timeout alone is not evidence of a bug and must not trigger an issue. Treat
+it as actionable only after tracing it to a PerlOnJava defect (for example, a
+reproducible infinite loop); retain a bounded reproducer that demonstrates the
+non-termination.
+
+## Investigation
+
+- Read the complete archived target log, especially configure output, the test
+  harness summary, and timeout context.
+- Check distribution metadata and source for XS/C code. Treat GUI/display,
+  native-library, network-service, and OS-only test requirements as
+  environment constraints until system Perl proves otherwise.
+- Run the focused upstream suite with system Perl and retain its full output.
+  Bound any potentially hanging invocation with `timeout`.
+- If the suite passes under system Perl, reproduce with both PerlOnJava
+  backends when the failure is compiler/runtime related.
+- When an internal PerlOnJava cause is confirmed, keep the smallest reusable
+  reproducer and record which backend(s) it affects. Do not infer an internal
+  cause solely from `Unknown test outcome`.
+- For timeouts, first rule out slow dependencies and environmental contention.
+  Create an issue only when a bounded reproducer attributes the timeout to a
+  PerlOnJava internal bug.
+- For XS modules, count or otherwise document reverse dependencies before
+  calling the module important. Prefer CPAN index/dependency metadata over a
+  name-based guess.
+
+## GitHub handling
+
+Search for an existing issue before proposing a new one. Use an existing
+`bug` label when appropriate; do not create a new label for an isolated CPAN
+failure. Include the distribution/version, system-Perl result, PerlOnJava
+result, environment prerequisites, and archived-log path in any issue.
+
+Do not create the issue unless the user explicitly asks for that external
+action.