Prove your AI agent works before you ship it.

A pack of slash commands and sub-agents for Claude Code and Cursor that turns “I think the agent works” into an eval suite, a trace review, a regression check, and a red-team pass — run from inside the editor you already use.

$29 one-time

Buy & download Watch 2-min demo

The problem

Agents fail quietly, and nobody notices until a user does

Testing an AI agent is not like testing a function. There is no single correct output, the failure modes are fuzzy, and every prompt change silently rewrites behaviour you already verified. Most teams end up eyeballing a handful of runs and calling it done — then find out in production that a model update broke a path nobody thought to check.

This toolkit gives that work a shape: written evals, a repeatable trace review, a regression check you run after every prompt or model change, and an adversarial pass to find the inputs you would not have thought of.

What's inside

Four commands, three sub-agents, both editors

/eval-suite — reads an agent spec and writes a full eval suite: cases, expected behaviour, scoring criteria.
/trace-review — walks an execution trace and points at where the agent drifted, not just that it did.
/regression-check — re-runs your verified behaviour after a prompt or model change and reports what moved.
/red-team — generates adversarial inputs aimed at the failure modes specific to your agent's job.
3 sub-agents — eval-writer, trace-analyst, regression-hunter, for the long-running work that shouldn't clog your main context.
Cursor rules — the same four workflows as .mdc rules, so the pack works in either editor.
A worked example + CLAUDE.md template — a real agent spec you can run the commands against on day one.

Claude Code Cursor 16 files No dependencies MIT-style license

How it runs

It lives in your repo, not in another dashboard

# drop the pack into your project
.claude/commands/   → 4 slash commands
.claude/agents/     → 3 sub-agents
.cursor/rules/      → same workflows for Cursor

# then, in Claude Code:
/eval-suite examples/example-agent-spec.md
/trace-review
/regression-check

Plain markdown and frontmatter. Nothing to install, no server to run, no account to create, no telemetry.

Fit

Who this is for — and who it isn't

A good fit if

You are building an agent or an LLM feature in Claude Code or Cursor, you change prompts or swap models often, and you want the QA step to be a command you run rather than a discipline you keep forgetting.

Not a good fit if

You need a hosted eval platform with dashboards, team seats, and stored run history — this is a local pack of prompts, deliberately. It also won't help much if your agent has no spec at all; write one first.

Details

Before you buy

What exactly do I get?: A zip of 16 source files — commands, sub-agents, Cursor rules, example, README, license. Instant download after checkout.
Do I need an API key or subscription?: No. It runs inside the Claude Code or Cursor setup you already pay for.
Updates?: Free for buyers. Updated files are pushed to the same download link, and you get an email when there's a new version.
Refunds and invoices?: Checkout, invoices, VAT and refunds are handled by Lemon Squeezy as merchant of record. If something isn't right, email me first — I'd rather fix it.

Get the toolkit

One-time $29. Download instantly, keep the files, use them in as many of your own projects as you like.

Buy & download — $29 Ask a question

AI-assisted, human-reviewed: every file was written and read line by line before release.