Conciseness

Mostly efficient: no section explains basic concepts at length, and the bulk is project-specific value (the exit-code mapping table, "JSON is the contract", determinism rules, the Bats 1.5+ `run --separate-stderr` detail). However, verbatim clig quotes like "Send output to `stdout`... Send messaging to `stderr`" and "Return zero exit code on success, non-zero on failure" restate rules Claude already knows. Not 5 because those redundant quotations could be trimmed to one-line citations; not 3 because no section is padded and every convention carries concrete non-obvious decisions.

Actionability

Mostly executable, copy-paste-ready guidance: `run mycli --nonexistent; [ "$status" -eq 2 ]`, `run --separate-stderr mycli list`, and `echo "$output" | jq -e '.[0] | has("name") and has("status")'` are real, runnable bats assertions. Not 5 because Convention 4's "Test pattern: golden-file comparison with sorted output" names the technique without showing a snippet, leaving one of the four in-body conventions without executable code; not 3 because the shown code is complete, correct, and covers the common cases.

Workflow Clarity

Each convention follows a clear repeated structure — rule, good/bad example, then test pattern — and the test patterns act as explicit checkpoints (assert the specific code, assert stream separation, assert the JSON schema). "Fix the output, not the assertion" gives unambiguous remediation direction, and the anti-patterns table maps each failure to a fix and convention. Not 5 because Convention 4's verification approach is named but not specified (no golden-file example); not 3 because the checkpoints for conventions 1-3 are explicit and the diagnostic path for flaky tests is clear.

Progressive Disclosure

Clean split, verified against the actual bundle: conventions 1-4 with test patterns stay inline (~200 lines), and "The remaining four conventions and their bats test patterns are in [references/conventions-5-8.md](references/conventions-5-8.md)" points to a real, one-level-deep file (75 lines) that contains its own organized sections and bats code. References section lists external resources without nesting. Not 4 because navigation is explicit, the split point is sensible, and there are no organization gaps or buried references.

Specificity

Enumerates multiple concrete capabilities with comprehensive coverage: "exit-code policy (0 success, non-zero failure with stable codes per failure mode)", "`stdout` for primary data / `stderr` for messages", "`--json` / `--plain` for machine-readable output", "deterministic ordering and timestamps", "`NO_COLOR` / TTY-aware color", "`-q` / `--verbose` discipline, and stable `--help` / `--version`" — every major CLI output surface is named, none vaguely. Not 4 because there are no coverage gaps; each named convention is a concrete, checkable behavior rather than a generic action.

Completeness

Both what and when are explicit and concrete: what is "Conventions for designing AND testing CLI output so it stays parseable and assertable" with the full convention enumeration, and when is "Use when designing or testing a CLI's output - a new flag, command, or error message, writing CLI assertions, or fixing flaky tests caused by non-deterministic output". Not 4 because the when-clause is fully explicit with concrete trigger phrases, not merely present or loosely worded.

Trigger Term Quality

Natural user phrasings appear on both the design and testing sides: "a new flag, command, or error message", "writing CLI assertions", "fixing flaky tests caused by non-deterministic output" — exactly what a user would say when they need this skill. Not 4 because it covers the natural synonyms across scenarios (designing output, writing tests, flaky/brittle tests, non-deterministic output, parseable/assertable) rather than leaving common variations out.

Distinctiveness Conflict Risk

Clear niche (CLI output contract) with active disambiguation from the closest overlapping skills: "it is the assertion contract for `bats-testing` and tells `tui-snapshot-tester` what does NOT need a snapshot". Not 4 because overlap risk is not just low but explicitly addressed by stating what this skill is and is not relative to its sister skills.