Agent skills · MIT · No account

Marketing skills for your coding agent.
It already has the repo.

Open-source skills for Claude Code, Codex, Cursor and any agent that reads SKILL.md. The first, launch-check, reads your live site and the code behind it, scores the launch out of 100 and ranks three fixes.

Install
npx skills add 5hirish/duct
Or as a plugin
Claude Code
claude plugin marketplace add 5hirish/duct && claude plugin install duct@duct
Codex
codex plugin marketplace add 5hirish/duct && codex plugin add duct@duct

Then ask your agent “Is my site ready to launch?”, or type /launch-check in Claude Code.

The first skill: launch-check

It reads the page. Then it reads the code.

01

Reads the site and the repo

Up to ten live pages, the sitemap and every card image, then the code behind them: the analytics tag, the signup or download event, where the head comes from, and the claims the page makes.

02

Scores six areas out of 100

Every finding names the URL or the file and line it came from. What it could not see is marked unverified, never failed.

03

Ranks three fixes, applies one on a yes

The three that matter most for launch day, each as a diff. It asks before every one and never commits, pushes or deploys.

What it checks

Six areas, weighted by what a launch loses.

A launch nobody can measure teaches you nothing, so measurement weighs most. A claim the code contradicts is the top comment on launch day, so claims come next.

Each area starts at 100. A FAIL costs 25 points and a WARN 10. Something it could not check costs nothing and is listed as unverified, so a missing repository lowers the confidence, not the score.

  • 85–100 Ready
  • 70–84 Ready after fixes
  • 55–69 Needs work
  • Below 55 Not ready
  1. 25Measurement

    Will the launch be recorded, down to the one action?

  2. 20Claims

    Does the product do what the page and the release notes say?

  3. 20Search

    Can search engines index and describe every key page?

  4. 15Social card

    Does a shared link show a title, a description and an image?

  5. 10Answer engines

    Can ChatGPT, Perplexity and Claude read and cite it?

  6. 10First impression

    Does the first screen say who it is for and what to do?

A real run

We ran it on getduct.ai. 96 out of 100.

getduct.ai and its repository, on November 10, 2026. Only two findings cleared the bar for a fix, and the first one exists only in the code: downloads from Europe would read low, with nothing on the page to show it.

Run it on yours
npx skills add 5hirish/duct
duct-launch-check.md

Launch check: getduct.ai

96/100, Ready · 10 pages read · repo: 5hirish/duct
The action: download the desktop app, tracked as download_started

Three fixes

Only two findings cleared the bar. The rest is polish.

  1. Count the downloads consent never sees.

    In the EEA, UK and Switzerland, analytics load only after Accept, and download_started waits in a queue that is never sent for a visitor who ignores the banner. Launch-day downloads from Europe will read low.

    Impact 2 · Effort 1 · Found by: repository · site/assets/duct.js:364

    Fix: no code. Read the release's download count from GitHub on launch day as the floor, and compare it with download_started.

  2. Stop telling every visitor the download is for macOS.

    The home page button says “Download for macOS” to Windows, Linux and phone visitors alike, because the per-platform label runs only on /download.

    Impact 2 · Effort 1 · Found by: both · site/index.html:308

    - Download for macOS ↓
    + Download Duct ↓
    Apply fix 2? (yes / no / skip all)

Post this

Ran a launch check on getduct.ai with my coding agent: 96/100. It found that downloads from Europe would undercount behind the cookie banner, which no page audit could have told me.

What it never does

No account. No upload. No change without a yes.

No account

No sign-up, no API key, no Duct login. It uses your agent's own web and shell tools, and never signs up, logs in or submits a form anywhere.

Nothing sent to Duct

The skill is instructions, not a service. Your site and your code are read by your agent and go nowhere else. It never opens .env or key files.

No change without a yes

It writes one file, its report. Every fix is a diff it shows first and applies only when you say yes. It never commits, pushes or deploys.

Where Duct comes in

The check is one moment. Duct reads the weeks after.

launch-check tells you whether the launch will be counted. Duct, the open-source AI marketing agent this shelf comes from, tells you what the count means.

Connect Duct to the same agent

Duct's MCP server lets the agent you just used list and read your Duct projects, and set one up from this repository. One address, one Allow in your browser. It can never approve a change.

Use Duct from your agent →

See what the launch did

The desktop app reads GA4, Search Console, your ads and Stripe together, answers why signups moved, and proposes the change, applied when you press Apply.

Download Duct ↓

FAQ

Before you install it.

Which agents does it work with?

Any agent the open skills CLI installs into: Claude Code, Codex, Cursor, OpenCode, Gemini CLI, GitHub Copilot, Windsurf and Zed among them. The skill is one SKILL.md folder in the Agent Skills format, so an agent that reads that format can use it without the installer. Claude Code and Codex can also install it as a plugin from Duct's own catalogue, with the commands under the install box.

Does it send my site or my code to Duct?

No. The skill is a set of instructions your agent follows with its own web and shell tools. There is no Duct server in the loop, no account and no API key. What your agent reads goes only where your agent's requests always go: to the model you already use.

Will it change my code?

Not without a yes. The one file it writes on its own is its report, duct-launch-check.md. Each fix comes as a diff and a question, one at a time. It never commits, pushes or deploys, and it never opens .env or key files.

Do I need a repository open?

No, but it is better with one. Without the code it checks the live pages only, and most of the measurement area comes back unverified rather than failed. The findings a page audit cannot give, like a download that sends no event, come from the repository.

How is it different from an SEO audit?

An SEO audit reads the HTML. launch-check also reads the code behind it and checks what the page promises against what the product does. Search is one area of six. For a full crawl with keyword clusters and competitors, run Duct's free SEO audit.

What does it cost, and how long does it take?

The skill is free and MIT licensed. It runs on your agent's own model, so a check costs what your agent spends reading ten pages and part of a repository. It stops at ten pages and aims to finish in ten minutes.

Does the installer track anything?

The skills CLI, which is open source and not ours, counts installs anonymously for the skills.sh leaderboard. Set DISABLE_TELEMETRY=1 or DO_NOT_TRACK=1 before you run it to turn that off. The skill itself sends nothing anywhere.

How do I update or remove it?

npx skills update brings it to the latest version, and npx skills remove launch-check takes it out. The install goes into the project you run it in; add -g to make it available in every project. The CLI needs Node.js 22.20 or later.

Which skills come next?

The ones people ask for. Tell us the marketing job you would hand your agent in GitHub Discussions.