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 ngocsangyem/MeowKit --skill evaluategit clone --depth 1 https://github.com/ngocsangyem/MeowKitWrote 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/ngocsangyem/meowkit/evaluate)<a href="https://agentmods.dev/skills/ngocsangyem/meowkit/evaluate"><img src="https://agentmods.dev/badge/skills/ngocsangyem/meowkit/evaluate/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/ngocsangyem/meowkit/evaluate"><img src="https://agentmods.dev/badge/skills/ngocsangyem/meowkit/evaluate.svg" alt="Reviewed on agentmods" width="80" height="20"></a>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 | $0.00035 | $0.01442 |
| Opus 5 | $0.00017 | $0.00721 |
| Sonnet 5 | $0.00007 | $0.00288 |
| Haiku 4.5 | $0.00003 | $0.00144 |
Grade A, and why
mk:evaluate scanned grade A with 1 finding 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 8d ago.
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.
Makes network callslowCapability
Not a fault in itself. Listed so you know the mod talks to something, and to what.
description: Experimental behavioral evaluator. Drives a running artifact via browser/curl/CLI and records rubric evidence; the runner is not yet a fully implemented automated evaluation system. How it starts
The opening of the file, as written. The whole thing — 117 lines — stays where its author put it; the contents beside it link to each section on GitHub.
mk:evaluate — Experimental Behavioral Verification
Step-file workflow that drives a running build, probes each rubric criterion via active verification, and produces a graded verdict with runtime evidence. Owned by the evaluator agent (Phase 3+).
When to Use
Activate when:
- User runs
/mk:evaluate <target>with a URL, file path, or running-app handle - A generator iteration completes and the harness needs a graded verdict
- After Phase 3 (build) and before Phase 5 (ship) for frontend/fullstack/CLI products
- When asked to "grade the running app", "check the build behaviorally", or "verify against the spec"
Skip when:
- The build has no runnable artifact (pure library, type-only package)
- The task is structural code review only — use
mk:reviewinstead - The task is
/mk:fixsimple — overhead exceeds value
Hard Constraints
- Active verification gate — every verdict MUST include non-empty
evidence/directory with at least one of: screenshot, HTTP response capture, CLI stdout+exit-code transcript.validate-verdict.shrejects PASS verdicts with empty evidence and converts them to FAIL. - Skeptic persona enforced — load
prompts/skeptic-persona.mdat session start. Re-anchor before each criterion grading. - Max 15 criteria per session — split into multiple sessions if rubric composition exceeds. Heuristic: context overflow risk above this threshold.
- No source code edits — evaluator owns
tasks/reviews/*-evalverdict.mdonly. Never modifies source files. - Frontend default preset is pruned —
frontend-apploads only product-depth, functionality, design-quality, originality (per Phase 2 v2.0.0 audit). Other 3 rubrics opt-in only.
Workflow
Execute via workflow.md. Step-file architecture — load one step at a time.
Step 1: Load Rubrics → compose preset, load skeptic persona, parse sprint contract if exists
Step 2: Boot App → start the build if not already running (skip if URL given)
Step 3: Probe Criteria → drive browser/curl/CLI per criterion; capture evidence
Step 4: Grade and Verdict → score each rubric; write verdict file; run validate-verdict.sh
Step 5: Generator Feedback → produce one-line fix guidance per FAIL/WARN; emit handoff message
What ships with it
11 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.
- prompts/skeptic-persona.md 5.9 KB
- references/active-verification-patterns.md 10 KB
- scripts/recompute-score.py 8.3 KB runs code
- scripts/run-evaluator.sh 4.5 KB runs code
- scripts/validate-verdict.sh 11 KB runs code
- step-01-load-rubrics.md 5.4 KB
- step-02-boot-app.md 3.7 KB
- step-03-probe-criteria.md 5.1 KB
- step-04-grade-and-verdict.md 6.7 KB
- step-05-feedback-to-generator.md 3.9 KB
- workflow.md 4.6 KB
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.
- 8d ago First seen · 117 lines · 35 tokens per session scan A fbea2c43cc45
mk:evaluate is a skill published in the GitHub repository ngocsangyem/MeowKit (14 stars, last pushed 1mo ago), licensed MIT. It adds 35 tokens to every session and 1,442 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 1 finding (makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-03.
Other skills, from other repositories
ui-test
Runs UI tests described in plain English by driving real Chrome via the Claude-in-Chrome extension. Covers end-to-end flows (clicks, forms, assertions), visual checks (screenshot + optional baseline diff), accessibility (axe-core), performance (Web Vitals + light Lighthouse-style metrics), and an interactive --debug…
audit-ui-e2e
Runs a beginner-mind end-to-end UI audit of any running app — local dev server, staging, production, or a specific URL. Drives Chrome through every interactive element on the target surface, collects structured findings (severity, category, where, symptom, impact, repro, triage), and hands the result off to…
capture-screens
Automatically navigates a web app using Playwright MCP and captures context-aware named screenshots at each product feature state. Names each file semantically based on context (e.g., checkout-payment-form-filled.png). Outputs a manifest.json mapping filenames to descriptions and a summary report. Use when documenting…
smoke-test
Traces and verifies that something works end-to-end in any environment. Builds a check plan from natural language input, confirms it, then runs each check reporting pass/fail. Use when validating deployments, pipelines, features, or migrations.
Playwright Test Builder
Generates robust Playwright end-to-end tests for web pages and user flows.
verify
How to run and drive this app to verify UI/server changes at runtime.