---
title: QA with Claude Code: install the skill, then ask in plain words
description: How Claude Code runs Kaloko: the skill it installs, what it may do without asking, what it asks first, the prompts that work, and how it reads reviewer feedback and the inbox.
seo_title: QA with Claude Code: the skill, prompts and canvas
seo_description: Install the kaloko skill in Claude Code, ask in plain words and read the result on the canvas: screens, criteria, evidence.
query: QA for Claude Code
kind: agent
order: 10
demo: true
---
Claude Code writes the feature anyway. With the `kaloko` skill it also walks the acceptance plan, captures the screens, evaluates the criteria, shares the canvas and later reads what the reviewer sent back.

## The problem

Without a skill the agent improvises QA: it opens the page once, says "looks fine" and moves on. Nobody can see what it looked at.

## What changes with Kaloko

The skill tells Claude Code when a walkthrough is due, what "done" means (validated scenario, every step captured or explained, evaluation and build run, the link posted), what it may do without asking (local and review environments, test data there, scenario mechanics) and what needs a person first (changing acceptance criteria, any production run, uploading).

## How it works

1. **Install**: `npm install --save-dev kaloko`, then `npx kaloko init --agent claude --org <your-org>` copies the skill to `.claude/skills/kaloko` and creates `kaloko.config.yml`. `npx kaloko doctor` checks Chrome, token, environments and the evaluator key.
2. **Ask**: "Walk the acceptance plan of TASK-123 on local and share the result." The skill pulls in the references it needs: acceptance, execution, evaluation, environments, collaboration.
3. **Drive the browser your way**: Claude Code uses the shared browser through a Playwright step script or a browser tool over MCP (the CDP endpoint is printed by `kaloko browser start`).
4. **Close the loop**: after sharing, the skill reads `kaloko feedback` and `kaloko inbox` before continuing on the task. Optionally connect the MCP server so the agent reads runs and feedback as tools.

## Skills and prompts

```
Walk the acceptance plan of TASK-123 with Kaloko on local and share the result into the PR.
```

```
What did the reviewer return on the last run of TASK-123? Fix it and run again.
```

```
Compare the latest run of checkout with the baseline and explain the differences.
```

Expected outcome in each case: the commands the skill prescribes, a summary line, failing or unreviewed criteria with evidence, and links to the canvas; no production run without asking.

## What you get

- A skill that encodes the rules once, so every task gets the same walkthrough.
- Runs, feedback and the inbox readable by the agent through the CLI or the MCP server at `kaloko.app/mcp`.
- A canvas the rest of the team can accept or return without touching the terminal.

## FAQ

### Does Claude Code need a browser tool?

No. Scripted steps run through Playwright inside the CLI; for judgement-driven steps a browser MCP or computer use helps, but `kaloko capture --url` works from the shell alone.

### What does the agent send to the AI evaluator?

For criteria marked `semantic`, a reduced text outline of the page with e-mails and tokens masked. Criteria marked `vision: true`, and text answers the evaluator is unsure about, also send a scaled screenshot (or the criterion's crop) to the vision evaluator: Anthropic with your `ANTHROPIC_API_KEY`, or through Kaloko. `evaluators: { vision: false }` in `kaloko.config.yml` turns that off. Otherwise screenshots and raw HTML go only to your own Kaloko organization.
