πŸ“¬ The AI growth newsletter - daily or weekly, your choice. + free 20-prompt pack.Subscribe β†’

Home β€Ί AI agent skills β€Ί Verification

Verification skills: which ones should your coding agent use?

13 skills help a coding agent check its own work. 9 come from the tool's own publisher and 3 can also push code or merge a PR.

None of them has been through our full review yet. Data read on October 31, 2026 from GitHub and each official page.

What each skill can do

Read in each skill's instructions and docs, without running them. An agent with terminal access can still do more than a skill describes.

SkillDrives a browserBrowserReads your codeReads codeWrites or postsWritesDeletesDeletesGoes onlineOnline
Official publishers
agent-browserOfficialin its instructionsnot seen in our readingnot seen in our readingnot seen in our readingin its instructions
Chrome DevTools MCPOfficialin its instructionsnot seen in our readingnot seen in our readingnot seen in our readingin its instructions
Code Review (Anthropic)Officialnot seen in our readingin its instructionsoptional or limitednot seen in our readingin its instructions
Differential Review (Trail of Bits)Officialnot seen in our readingin its instructionsoptional or limitednot seen in our readingoptional or limited
Playwright CLIOfficialin its instructionsoptional or limitedoptional or limitednot seen in our readingin its instructions
Playwright MCPOfficialin its instructionsnot seen in our readingnot seen in our readingnot seen in our readingin its instructions
Security Review (Sentry)Officialnot seen in our readingin its instructionsnot seen in our readingnot seen in our readingnot seen in our reading
ThermosOfficialnot seen in our readingin its instructionsnot seen in our readingnot seen in our readingoptional or limited
Verification (Vercel)Officialin its instructionsin its instructionsnot seen in our readingnot seen in our readingin its instructions
Other authors
Cursor Team KitCursor repoin its instructionsin its instructionsin its instructionsoptional or limitedin its instructions
Playwright SkillCommunityin its instructionsoptional or limitedoptional or limitednot seen in our readingin its instructions
pstackCursor repoin its instructionsin its instructionsin its instructionsin its instructionsin its instructions
SuperpowersCommunitynot seen in our readingin its instructionsin its instructionsin its instructionsin its instructions

in its instructions optional or limited not seen in our reading

Check before installing 3

Their instructions include steps that push code, merge a PR or delete branches. Read them and keep an eye on what the agent does.

pstack

Can merge PRs and force push.

The 13 skills

Official publishers first, then A to Z. It is not a ranking. 13 of 13 had a commit in the last 6 months.

What to check
Publisher
Skill, Official

agent-browser

Active Not reviewed yet
Checks
Web flows in a real browser, from the terminal.
Permissions
No browsing limit by default.
Activity
todayOct 31, 2026, Apache 2.0
Details, install and sources

Rust command line tool that drives a browser for AI agents, with a skill to install. The agent opens pages, clicks, fills in forms and reads the result.

Security note (passive reading) : By default, no limits on navigation, actions or output. An allowed domains list (--allowed-domains) is optional. Installs a native binary through npm.

Install : npm install -g agent-browser then npx skills add vercel-labs/agent-browser

Works in : Claude Code, Codex, Cursor, Gemini CLI

GitHub stars : 45k

Sources : 1

MCP server, Official

Chrome DevTools MCP

Active Not reviewed yet
Checks
Performance, console errors, network and accessibility of a page.
Permissions
Usage statistics sent to Google by default.
Activity
yesterdayOct 30, 2026, Apache 2.0
Details, install and sources

MCP server and CLI that let the agent control and inspect a live Chrome: performance traces, network requests, console, screenshots. Ships with skills (accessibility, memory leaks, LCP).

Security note (passive reading) : Google collects usage statistics by default (turn off with --no-usage-statistics). Performance tools may send trace URLs to Google's CrUX API (--no-performance-crux to avoid it).

Install : npx -y chrome-devtools-mcp@latest (in the MCP config)

Works in : MCP

GitHub stars : 55k

Sources : 1

Plugin, Official

Code Review (Anthropic)

Active Not reviewed yet
Checks
Review of a pull request before merge.
Permissions
Can post a comment on your PR.
Activity
5 days agoOct 26, 2026 (folder), Apache 2.0
Details, install and sources

Official Claude Code plugin: the /code-review command runs 4 agents in parallel on a pull request (CLAUDE.md rules, bugs, git history), scores each issue from 0 to 100 and keeps only those at 80 or more.

Security note (passive reading) : The command allows read gh commands (pr diff, pr view, issue list...) and gh pr comment. With the --comment option, the review is posted on the PR, visible to others. Needs gh signed in to your GitHub account.

Install : Claude Code's official marketplace (claude-plugins-official)

Works in : Claude Code

GitHub stars : 45k (whole claude-plugins-official repo)

Sources : 1, 2

Active Not reviewed yet
Checks
Security review of a change, with git history and untested areas.
Permissions
Writes a report file in your project.
Activity
45 days agoSep 16, 2026 (folder), CC BY-SA 4.0
Details, install and sources

Plugin by Trail of Bits, a security audit firm. It reviews a change (PR, commit or diff) for security: git history of removed code, number of callers affected, changed code with no test, then a markdown report.

Security note (passive reading) : The skill allows Read, Write, Grep, Glob and Bash. It always writes a report file in your project. For a PR, it reads its details with gh. CC BY-SA 4.0 licence, a content licence rather than a software one.

Install : /plugin marketplace add trailofbits/skills then /plugin menu (Claude Code)

Works in : Claude Code, Codex

GitHub stars : 7.9k (whole trailofbits/skills repo)

Sources : 1, 2, 3, 4

Skill, Official

Playwright CLI

Active Not reviewed yet
Checks
Flows in a real browser, screenshots, Playwright tests.
Permissions
playwright-cli commands only, open browsing.
Activity
todayOct 31, 2026, Apache 2.0
Details, install and sources

Playwright command line tool, shipped with skills for coding agents. The agent opens a browser, tests pages, takes screenshots and works with Playwright tests.

Security note (passive reading) : The skill allows the playwright-cli and npx playwright commands. Installs a global npm package. The driven browser can open any site.

Install : npm install -g @playwright/cli@latest then playwright-cli install --skills

Works in : Claude Code

GitHub stars : 15k

Sources : 1, 2

MCP server, Official

Playwright MCP

Active Not reviewed yet
Checks
Clicks, forms and page content in a real browser.
Permissions
Open browsing. Not a security boundary, per its docs.
Activity
todayOct 31, 2026, Apache 2.0
Details, install and sources

MCP server that gives the agent a browser driven by Playwright. The agent reads the page through the accessibility tree, without screenshots or a vision model.

Security note (passive reading) : The documentation states that Playwright MCP is not a security boundary, and that allowed or blocked origin lists are not one either. Optional capabilities: vision, pdf, devtools.

Install : claude mcp add playwright npx @playwright/mcp@latest

Works in : MCP

GitHub stars : 39k

Sources : 1

Active Not reviewed yet
Checks
Security review of code, reporting only confirmed flaws.
Permissions
No write tool in the skill.
Activity
12 days agoOct 19, 2026 (folder), Apache 2.0 (OWASP references under CC BY-SA 4.0)
Details, install and sources

Skill by the Sentry team, from its public repo of internal skills. It reviews code for flaws (injection, XSS, authentication, cryptography) and only reports what it confirmed by tracing data through the code.

Security note (passive reading) : The skill allows Read, Grep, Glob, Bash and Task, with no write tool. Installing with npx skills add or the plugin also adds the repo's other skills, some of which commit or open PRs.

Install : npx skills add getsentry/skills (installs the whole repo)

Works in : Claude Code, Cursor, Cline, GitHub Copilot

GitHub stars : 1.1k (whole getsentry/skills repo)

Sources : 1, 2, 3

Plugin, Official

Thermos

Active Not reviewed yet
Checks
Double review of a branch: bugs and security on one side, code quality on the other.
Permissions
Read only per the skills we read, called by name.
Activity
3 months agoAug 8, 2026 (folder), MIT
Details, install and sources

Cursor's review plugin. It runs two reviews in parallel on the same branch: one looks for bugs, regressions and security flaws, the other judges code quality and structure. The agent then merges both reports.

Security note (passive reading) : The three skills only run when you call them by name (disable-model-invocation). They read the diff and the changed files and return a report. The security review may read the PR discussion with gh or glab. The plugin description mentions "take-the-wheel" and "FSD merge-ready" flows that we did not find in the folder.

Install : /add-plugin thermos (Cursor)

Works in : Cursor

GitHub stars : 15k (whole cursor/plugins repo)

Sources : 1, 2, 3, 4

Plugin, Official

Verification (Vercel)

Active Not reviewed yet
Checks
A feature's whole flow: browser, API, data, response.
Permissions
Reads .env.local, telemetry on by default.
Activity
25 days agoOct 6, 2026 (folder), Apache 2.0
Details, install and sources

Skill from Vercel's official plugin. It infers the feature you are building, then checks the whole flow: page in the browser, API route, data, response. It stops at the first broken point and reports it with evidence.

Security note (passive reading) : The skill reads .env.local files, runs vercel logs and vercel env ls, and opens the page in a browser for the console and screenshots. It is part of a 37-skill plugin with session hooks. Plugin telemetry is on by default (VERCEL_PLUGIN_TELEMETRY=off turns it off).

Install : npx plugins add vercel/vercel-plugin (installs the whole plugin)

Works in : Claude Code, Cursor, Codex, GitHub Copilot, Grok Build, OpenCode

GitHub stars : 321

Sources : 1, 2, 3

Plugin, Cursor repo

Cursor Team Kit

Active Not reviewed yet
Checks
Before and after evidence for a fix, Playwright tests, UI driven in a browser, build checks.
Permissions
Some skills commit, push and open PRs.
Activity
53 days agoSep 8, 2026 (folder), MIT
Details, install and sources

Plugin published in Cursor's plugins repo, by Eric Zakariasson: 18 everyday workflow skills. For checking work: verify-this proves or disproves a claim with before and after evidence, run-smoke-tests runs Playwright tests, control-ui drives your UI in a browser, check-compiler-errors runs the build.

Security note (passive reading) : Several skills in the kit act on git and GitHub: review-and-ship and new-branch-and-pr commit, push the branch and open a PR, loop-on-ci pushes fixes until CI passes. make-pr-easy-to-review can rewrite history after proposing a plan. These skills can trigger on their own, except two that need to be called by name.

Install : /add-plugin cursor-team-kit (Cursor)

Works in : Cursor

GitHub stars : 15k (whole cursor/plugins repo)

Sources : 1, 2, 3, 4

Skill, Community

Playwright Skill

Active Not reviewed yet
Checks
Custom Playwright scripts, including your local pages.
Permissions
Runs node and npm, reads and writes files.
Activity
3 months agoAug 14, 2026, MIT
Details, install and sources

Skill that has the agent write and run Playwright scripts: page tests, multi-step flows, screenshots, network interception. Detects local dev servers.

Security note (passive reading) : The skill allows node, npm, and reading and writing files. First setup downloads Playwright and Chromium. Scripts are written to the temp folder, then run with node. Visible browser by default.

Install : npx skills add lackeyjb/playwright-skill --skill playwright-skill --global

Works in : Claude Code

GitHub stars : 3.3k

Sources : 1, 2

Plugin, Cursor repo

pstack

Active Not reviewed yet
Checks
Launches the app, drives it and keeps evidence (screenshots, logs); adversarial multi-model review; tests written before the fix.
Permissions
Can merge PRs and force push.
Activity
3 days agoOct 28, 2026 (folder), MIT
Details, install and sources

Plugin by Lauren Tan, engineer at Cursor: 58 skills and working principles. Several are about checking work: create-verification-skill generates a skill that launches your app and drives it like a user, interrogate has several models review a change, tdd writes a failing test first.

Security note (passive reading) : The playbooks tell the agent to run git and gh, up to merging PRs (gh pr merge) and pushing with --force-with-lease. The folder holds 63 script files. One document shows a curl | sudo sh install of a third-party tool. The verification skills we read only run when you call them by name.

Install : /add-plugin pstack (Cursor); /plugin install pstack@cursor-plugins (Claude Code)

Works in : Cursor, Claude Code

GitHub stars : 15k (whole cursor/plugins repo)

Sources : 1, 2, 3

Plugin, Community

Superpowers

Active Not reviewed yet
Checks
Tests written before the code, verification commands rerun before saying "done", code review.
Permissions
Loads every session, can merge and push.
Activity
10 days agoOct 21, 2026, MIT
Details, install and sources

Development method made of skills for coding agents. On the checking side: test-driven-development, verification-before-completion (no claim of success without fresh evidence), requesting and receiving code review, systematic debugging.

Security note (passive reading) : A session start hook loads the skills automatically. The optional visual companion loads a logo from the maker's site with the version number (turn off with SUPERPOWERS_DISABLE_TELEMETRY).

Install : /plugin install superpowers@claude-plugins-official (Claude Code)

Works in : Claude Code, Codex, Cursor, Gemini CLI, Grok Build

GitHub stars : 310k

Sources : 1, 2

Method, sources and limits
  • Order: official publisher first, then A to Z. It is not a ranking, and affiliate links never change it. None of these entries has an affiliate link.
  • "Official" means the skill is published by the company behind the tool or the repo (Microsoft, Google, Vercel, Anthropic, Cursor, Trail of Bits, Sentry). pstack and Cursor Team Kit sit in Cursor's plugins repo but are signed by individual authors, so we list them as "Cursor repo".
  • Activity: date of the last commit read with git and the GitHub API on October 31, 2026 (Paris time). When a skill sits in a folder of a larger repo, it is the last commit on that folder.
  • Stars: GitHub REST API, October 31, 2026. When a skill sits in a larger repo, the count is for the whole repo.
  • The permissions table and security notes come from a passive reading of the README, SKILL.md files and plugin manifests. We did not run these skills. "Not reviewed yet" means our full review method has not been applied. Few marks in the table does not mean a skill carries no risk.
  • Not included: Anthropic's webapp-testing skill and Sentry's find-bugs and code-review skills (last commit on their folder in April 2026, more than 6 months ago), the third-party pstack mirror backnotprop/pstack, community ports of Thermos, and OpenAI's playwright skill, which wraps Playwright CLI already listed here.

🎁 Get the pack: 20 AI prompts to win clients

Free and available right away. Then email tips to pick AI tools for your goal. One-click unsubscribe.

↑