Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add GoogleCloudPlatform/evalbench --skill evalbench-reviewgit clone --depth 1 https://github.com/GoogleCloudPlatform/evalbenchWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/googlecloudplatform/evalbench/evalbench-review)<a href="https://agentmods.dev/skills/googlecloudplatform/evalbench/evalbench-review"><img src="https://agentmods.dev/badge/skills/googlecloudplatform/evalbench/evalbench-review/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/googlecloudplatform/evalbench/evalbench-review"><img src="https://agentmods.dev/badge/skills/googlecloudplatform/evalbench/evalbench-review.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector warn
SkillSpector: 1 finding, up to medium
These are SkillSpector’s own severities. On a checked sample its high-severity flags on skills were ~96% false positives — a documented command, a public API, a “never do X” rule — so we show them as a caution to read, not a verdict. Why →
- medium Excessive Agency · line 119 Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.Fix: Add human-in-the-loop confirmation for destructive, irreversible, or high-impact operations. Never auto-execute commands that modify files, send data, or alter system state.
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00170 | $0.03984 |
| Opus 5.5 | $0.00068 | $0.01594 |
| Sonnet 5.5 | $0.00034 | $0.00797 |
| Haiku 4.5 | $0.00017 | $0.00398 |
Grade A, and why
evalbench-review scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 287 lines — stays where its author put it; the contents beside it link to each section on GitHub.
EvalBench code review
Review a change on four axes, in this order. Verification comes first — an architecturally beautiful change that doesn't run is still broken, and running the code surfaces real defects that reading alone will not.
- Does it work? Prove it by execution, not by reading.
- Does it follow EvalBench architecture? Registration wiring, base-class contracts, import style, isolation, concurrency, config-driven behavior, docs.
- Does it still ship? The image builds, and GKE / Cloud Run / Cloud Build still work. A green local suite says nothing about any of these.
- Is it a good PR? Reviewable size, single purpose, a description that answers what and why, and code that is clear, simple, concise, maintainable and idiomatic.
Do not skip phase 1 because a diff "looks obviously fine". Do not stop after phase 1 because the tests pass — a change can pass every test and still be wired into the wrong layer, break the container build, or be unreviewable.
Phase 0 — Scope the diff
Establish exactly what changed before reading anything else.
BASE=$(git merge-base HEAD origin/main)
git diff --stat $BASE...HEAD
git diff -M $BASE...HEAD # -M so renames don't read as rewrites
git status --short # uncommitted work counts as part of the change
For a PR under review: gh pr diff <n> and gh pr view <n>.
Size the change first — it sets how you review everything else
Defect detection degrades as diffs grow, so measure before you read:
git diff --shortstat $BASE...HEAD -- . \
':(exclude)uv.lock' ':(exclude)CHANGELOG.md' \
':(exclude)evalbench/evalproto/*_pb2*' ':(exclude)datasets/**/*.json'
Generated and vendored files are volume, not review surface — uv.lock alone
is 5,000 lines. Judge against the hand-written count:
- Under ~400 lines — review normally.
- Over ~400 lines — report it as a
pr-hygienefinding with a concrete split proposal, naming the seams. - Any size — pace at no more than ~500 lines per hour. If the change is larger than you can review at that rate, say which parts you read carefully and which you skimmed. Never imply uniform depth you didn't apply.
What ships with it
4 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday Changed b7682b466368
- 1mo ago First seen · 287 lines · 170 tokens per session scan A 3ea17af0b31e
evalbench-review is a skill published in the GitHub repository GoogleCloudPlatform/evalbench (59 stars, last pushed today), licensed Apache-2.0. It adds 170 tokens to every session and 3,984 once invoked, about $0.0007 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
Eval-Driven Skill Design
A review guide for new coding-agent skills that checks whether each skill has both examples it should catch and examples it should ignore, plus evaluation setup.
adk-go-self-review
Review an ADK Go change the way a maintainer will — a fresh-context pass over the whole diff, five lenses (correctness and tests, scope, simplicity, style, adk-python parity), and the mutation check that proves your tests pin the change. Use before opening a PR, before any later push that changes code, and when asked…
adversarial-reviewer
Stress-test a code change for concrete correctness defects, unsafe assumptions, and failure modes. Use when the user explicitly asks for an adversarial review, a deep bug hunt, or to tear a change apart. Do not use for ordinary review, style feedback, or general improvement suggestions.
comet-verify
Comet Phase 4: Verify and Close. Invoke with /comet-verify. Verify implementation matches design, handle development branch.
go-testing
Trigger: Go tests, go test coverage, Bubbletea teatest, golden files. Apply focused Go testing patterns.
semgrep-rule-variant-creator
Creates language variants of existing Semgrep rules. Use when porting a Semgrep rule to specified target languages. Takes an existing rule and target languages as input, produces independent rule+test directories for each language.