Confluent Skill Reviewer

confluent-skill-reviewer — audit a Confluent agent skill

Three bars every skill in this repo must clear:

1. Agent Skills spec — frontmatter shape, naming rules, directory layout, token budgets (https://agentskills.io/specification).
2. Confluent conventions — lazy-loaded references, anti-trigger clauses, mode-table branching, ≥90% evals, SME + DTX/DevRel sign-off (CLAUDE.md, .github/pull_request_template.md).
3. Evals-as-contract — evals/evals.json with specific, verifiable expectations; fixtures kept in sync.

This skill walks an agent through a structured audit and returns a single Markdown report. It is read-only by default — never modifies the skill under review.

Pre-flight gates

Before producing any findings, confirm all three out loud (briefly):

1. Scope: which mode are you in? See Mode Detection. If unclear, ask the user once.
2. Target paths exist: list the skills/<name>/ directories you will audit. Stop if none.
3. Tool availability: run bash skills/confluent-skill-reviewer/scripts/run_skill_validator.sh --probe (from the repo root) to check whether the external skill-validator binary is installed. If absent, note it in the report and continue with native checks — do not block.

Skipping these gates is the most common source of bad reviews. The point is to be explicit about scope so the user can redirect early.

Mode detection

If you cannot tell, ask the user once: "Are you reviewing a PR diff, a single skill, or the whole repo?" Then proceed.

Check phases

Run phases in order. Each phase reads its reference only if a finding fires. Collect findings into one in-memory list with severity tags:

• Blocking — violates the spec, CLAUDE.md, or a PR-template gate. Must fix before merge.
• Warning — violates a convention but won't break tooling; reviewer judgment.
• Nit — style/clarity, no functional impact.

Phase A — Structural & spec conformance

Run bash skills/confluent-skill-reviewer/scripts/run_skill_validator.sh <skill-path> (from the repo root). Three outcomes:

• Binary installed → script emits skill-validator's JSON. Parse results[].level == "error" into Blocking, "warning" into Warning. Map each finding's file and line into the report.
• Binary missing → script exits 0 with an install hint on stderr. Note "skill-validator not installed, skipping spec checks" as a Warning in the report and do the spec checks natively (read references/spec-conformance.md for the rule list and walk through them).
• Binary present but the skill is multi-skill (path lacks SKILL.md) → script falls through to the per-skill JSON; treat each entry independently.

Then, regardless of validator state, native checks that the validator does not cover well in this repo:

• The skill's directory name matches the name: frontmatter field exactly.
• evals/ is not flagged as orphan (this repo expects evals; pass --allow-dirs=evals if invoking the validator directly).

Read references/spec-conformance.md only when interpreting an unfamiliar finding code.

Phase B — Confluent conventions

Inspect the SKILL.md against rules in CLAUDE.md § Skill anatomy and § Evals are the contract. The high-leverage checks:

1. Lazy-loading: does the SKILL.md inline the contents of any file under references/? Grep for headings that also appear in references and for long fenced code blocks that duplicate reference material. Inlined reference content is Blocking.
2. Anti-trigger clause: does the description: contain a Do NOT trigger for… clause? Absence is Blocking when neighbor descriptions share keywords (Phase C confirms); otherwise Warning.
3. Mode-table branching: if SKILL.md exceeds ~200 lines or covers more than one distinct workflow (e.g. build and debug), expect a mode-detection table near the top. Absence is Warning.
4. Reference depth: references nested more than one level below the skill root are Blocking (the spec restricts this).
5. Provenance metadata: skills authored in this repo declare metadata.author: confluent, metadata.version (semver), and metadata.last_updated (YYYY-MM-DD). Missing fields are a Warning. A compatibility field is expected when the skill requires specific packages, CLI tools, or environment access — absence on a skill that obviously has env requirements is a Nit.
6. Platform scoping: if the name contains a platform token (confluent-cloud-*, confluent-platform-*, warpstream-*, apache-kafka-*), the body must scope its instructions to that platform — Blocking if it generically covers all platforms. Conversely, if the name lacks a platform token but the body performs platform-divergent operations (Cloud API keys, WarpStream object-storage config, on-prem SASL), expect a platform-detection step and per-platform reference files (references/confluent-cloud.md, references/warpstream.md, etc.). Missing either is a Warning.
7. Plan-before-execute: skills that create, modify, or delete resources (Kafka topics, schemas, Flink statements, connectors, Terraform state) must include an explicit "present the plan, wait for user confirmation" step before any resource-modifying call. Absence in a CRUD-capable skill is Blocking.
8. Credential handling: skills that read credentials must not cat, Read, head, or grep a .env file directly — they should reference variables by name ($BOOTSTRAP_SERVERS) and verify presence with test -n "$VAR". SKILL.md or bundled scripts that read .env contents are Blocking. A credential-consuming skill with no guardrail language at all is a Warning.
9. Synthetic data only: run python3 skills/confluent-skill-reviewer/scripts/scan_pii.py <skill-path> (from the repo root) to catch real customer data or secrets committed into the skill — SKILL.md, references/, eval prompts, on-disk fixtures, and bundled scripts must use synthetic data only. The script emits JSON; map each finding into the report. severity: "blocking" (US SSN, Luhn-valid payment card, AWS key id, private-key block) → Blocking; severity: "warning" (email on a non-example domain, phone number) → Warning. CI runs the same scanner over skills/, so a clean run here mirrors the gate.

Read references/confluent-conventions.md for the full rule list and PR-template gates.

Phase C — Trigger overlap

Run python3 skills/confluent-skill-reviewer/scripts/check_trigger_overlap.py <root> (from the repo root). The script accepts either a repo root (it scans <root>/skills/*/SKILL.md) or a "skills root" directory (scans <root>/*/SKILL.md) — use the latter for evals/mock-skills/ runs. It parses each SKILL.md frontmatter, tokenises the description: field (filtering stopwords and domain-broad terms like confluent, kafka, schema, producer, consumer, topic, stream), and reports keyword collisions. For each collision:

• ≥3 overlapping non-broad keywords (e.g. "topology", "rebalancing", "windowing") with no mutual anti-triggers naming each other → Blocking.
• 2 overlapping non-broad keywords without mutual anti-triggers → Warning.
• Single non-broad keyword overlap → Nit.
• Overlap entirely on filtered domain-broad terms → script silently passes (these are coincidence, not collisions).

Read references/trigger-overlap.md only when proposing the wording of an anti-trigger fix — it has worked examples drawn from this repo's existing skills.

Phase D — Evals contract

Run python3 skills/confluent-skill-reviewer/scripts/check_eval_schema.py <skill-path>/evals/evals.json (from the repo root). The script validates:

• Top-level skill_name (string) and evals (array) keys present.
• Each eval has id, prompt, expected_output, files, and its checks under the assertions key. The repo standardized on one shape: assertions holding a list of strings. Two deviations are now Blocking — the legacy expectations key (must be renamed to assertions) and object-form entries [{id, type, description, …}] (must be flattened to strings).
• prompt is realistic user phrasing, not abstract (heuristic: ≥40 chars, not just "Build me an X"). Short prompts are a Warning.
• assertions[] are specific (heuristic: contain a verb, a noun, and at least one concrete identifier — file path, class name, config key, or "NOT" clause). Vague assertions are a Warning; cite CLAUDE.md § Evals are the contract: "expectations encode hard-won correctness — treat them as regression tests, not aspirations".

Cross-check fixture sync: each entry in files: [...] is either an on-disk path (relative to the skill root — must resolve when the skill has an evals/mock-repos/ or evals/mock-skills/ directory; missing fixtures are Blocking) or an inline fixture object {path, content} that carries the file body in the eval itself (not resolved on disk).

Read references/evals-contract.md for the JSON schema, both expectation shapes, and worked examples of weak vs strong expectations.

Phase E — PR template gates (PR-diff mode only)

Skip in single-skill and repo-wide modes.

Read .github/pull_request_template.md once for the live checklist. For the current PR, verify each item:

Read references/confluent-conventions.md § PR template if the gate wording in the live template has drifted.

Report format

Emit one Markdown report. Group findings by severity, not by phase — reviewers scan top-down:

# Skill review: <skill name or PR title>

**Mode:** PR-diff | Single-skill | Repo-wide
**Validator:** installed (v1.5.6) | not installed (spec checks done natively)
**Scope:** <list of skills audited>

## Blocking (N)

- `skills/<name>/SKILL.md:14` — Inlined contents of `references/build-templates.md` into the SKILL.md body. CLAUDE.md § Skill anatomy (lazy-load references bullet) requires lazy-loaded references. Move the content back and route to it from a decision point.
- `skills/<name>/evals/evals.json:42` — Eval id 3 references `evals/mock-repos/missing/` which does not exist on disk. Add the fixture or update the path.

## PR-template checklist (PR-diff mode only)

- [x] Docs updated
- [ ] Evals at 90%+ — author has not pasted score
- [x] SME reviewer assigned (@alice)
- [ ] DTX/DevRel reviewer — none assigned

## Summary

Recommend: **request changes** | **approve with nits** | **approve**.

If there are zero Blocking and zero Warning findings, the recommendation is approve. One or more Blocking → request changes.

What this skill does NOT do

• Does not edit the skill under review. Findings only.
• Does not run the skill's own evals (the harness lives in ~/.claude/skills/skill-creator/; the agent should invoke it separately if they want a 90% score).
• Does not analyze runtime/log behavior — that's tools/skill_review_dashboard.py.
• Does not review files outside skills/ (the build system, Makefile, service.yml, and .claude-plugin/ are out of scope; flag those for a human reviewer if changed).