Produck Skills: four markdown skills, one behind a token

Four skill folders, and only one of them needs a network service

The collection is four skill folders under skills/, short enough to take in at once. user-alignment turns vague or messy user requests into PRDs carrying scope, bounded phases, testable acceptance criteria and explicit do-not-do boundaries. produck-feedback pairs with it and drives a loop from a feedback ticket to an aligned PRD to a build. demo-studio turns a written brief into an animated launch video. taste runs a screenshot-driven pass over hierarchy, trust, spacing, CTA and consistency.

The split that matters for an adopter is that only one of the four reaches a network service. The other three are folders of markdown that a coding agent reads and follows. produck-feedback calls the Produck MCP, a hosted server, so that skill stays inert until you mint a token. One install line therefore gets you three working skills immediately and a fourth only after you sign up at tryproduck.com/app.

One npx line, and why nothing shows up in the session you have open

Installation is a single line:

npx skills add tryproduck/produck-skills

It runs through the open skills CLI from vercel-labs/skills, needs no global install, and the project claims support across 70+ agents including Claude Code, Cursor, Codex and Copilot. The files land where your agent expects to find skills, and they are available the next time that agent starts rather than in a session already running.

Understand what arrives. Nothing in this repository executes on its own. You get markdown instructions, plus .claude-plugin/ and .codex-plugin/ directories and a .mcp.json at the top level that let Claude Code and Codex treat the repository as a plugin marketplace. Behaviour changes because a model reads those files, which is why editing the text is a functional change rather than a cosmetic one.

What search_feedback and get_feedback return, and the token in front of them

produck-feedback depends on two tools exposed by the hosted Produck MCP server. search_feedback finds feedback tickets relevant to a piece of work and can filter by domain. get_feedback loads a single ticket in full: the user's written or spoken feedback, reconstructed page HTML, annotation markers, session activity and a generated design doc.

For Claude Code the token goes into a file the plugin reads at runtime.

# 1) in your shell: save your token
mkdir -p ~/.produck && echo "mcp_live_YOUR_TOKEN" > ~/.produck/mcp_token

# 2) in Claude Code:
/plugin marketplace add tryproduck/produck-skills
/plugin install produck@produck

Codex takes the same credential as an environment variable.

# 1) in your shell: export your token
export PRODUCK_MCP_TOKEN=mcp_live_YOUR_TOKEN

# 2) add the marketplace, then enable "produck" from the /plugins browser
codex plugin marketplace add tryproduck/produck-skills

You mint the token at tryproduck.com/app under MCP, where it takes the shape mcp_live_.... The consequence for anyone whose feedback lives elsewhere is direct: these skills cannot be pointed at a self-hosted equivalent, because the reconstructed page HTML and session activity arrive from the vendor's service. The half of the collection that closes the feedback loop only closes it over Produck's data.

The taste benchmark is the authors grading their own rubric on a fictional page

The most quotable evidence in the repository is also the least independent. It describes a one-shot website improvement benchmark run on the same intentionally flawed fictional SaaS landing page, twice: once as a baseline with feature-led platform copy, generic dashboard UI, weak proof and a noisy hierarchy, and once with the taste skill's category thesis, worldview sequence, domain artifacts, rubric and screenshot verification loop. Screenshots were captured at 1440 x 1100 with npx -y chrome-devtools-axi and combined into one before and after asset. Across nine axes on a one to five scale the baseline scores 1, 1, 1, 2, 1, 1, 2, 1, 1 and the taste-guided page scores 5, 5, 5, 5, 4, 5, 5, 5, 5. The improved page is built around a single category idea, the judgment layer for product feedback, opens with the claim that feedback is abundant while judgment is scarce, and swaps fake dashboard polish for opportunity memos, accepted and rejected evidence, and decision briefs.

The inputs sit in examples/taste-benchmark, and the rubric that produced those scores sits in skills/taste/references/taste-principles.md. One repository ships the skill and grades it, on one invented page, in one run, with the scoring instrument checked into the same tree. Read the table as a worked example of the loop the skill asks you to run on your own page. It is not evidence that your landing page will move those numbers.

demo-studio renders a film of an invented interface, not of your product

demo-studio asks for a written brief and returns a scene-by-scene animated video with motion graphics, camera moves, real UI mockups and an original sound kit, rendered from HTML and CSS with GSAP to MP4 through HyperFrames. It ships a runnable example film and a 25-sound synthesized SFX kit.

Two limits deserve naming before anyone schedules a launch video. HyperFrames is not in this repository: the top-level entries are .claude-plugin/, .codex-plugin/, .gitignore, .mcp.json, CONTRIBUTING.md, LICENSE, README.md, docs/, examples/ and skills/, with examples/taste-benchmark/ as the only example directory. The renderer is external, so the skill inherits whatever that dependency does and whatever it costs, and the repository's declared primary language of HTML describes the authoring medium rather than anything you run. Second, real UI mockups means the interface on screen was drawn by the agent to look convincing. The output is not a recording of your application, so it cannot stand in for one in a launch post where viewers will compare the film against what they actually get.

No releases, so the install has no version to resolve to

There are no GitHub releases in this repository, which is the practical fact about upgrading it. The project is not archived and the last push was on 2026-08-14, so nothing here suggests a dead repository, but npx skills add tryproduck/produck-skills has no version to pin to. You install whatever main holds at that moment, and a later run gives you something different with no changelog to read.

For four folders that may be tolerable. For instructions your agent reads on every run, it means a text edit lands as a behaviour change inside your workflow with no version boundary to roll back to. CONTRIBUTING.md documents the authoring contract, the folder layout, required frontmatter and a PR checklist, and states that the bar is high-quality, UX-centered skills only. That is governance about quality, not compatibility promises between revisions. The licence is Apache-2.0, which covers the markdown, so vendoring these files into your own repository and rewriting the wording is unencumbered.

The commercial centre, and where your feedback content goes

These skills come from the team behind Produck, a hosted product for capturing in-context user feedback, with documentation at docs.tryproduck.com and the site at tryproduck.com. That shapes what you can adopt and what you cannot. The Apache-2.0 grant covers the skill files, so forking them and editing the wording carries no licensing obstacle. It says nothing about the hosted MCP server or about the ticket content that passes through it, which is material your users typed and spoke. If your organisation has rules about where user feedback may be stored, those rules land on the vendor's service rather than on this repository, and the README does not document retention, access control or deletion for ticket content once get_feedback has fetched it.

The four skills are also uneven in how far you can evaluate them before committing. user-alignment and taste are text you can read and disagree with before installing anything. produck-feedback cannot be judged at all until the token exists and a live ticket comes back.

Keeping the same guidance inside your own repository instead

The alternative to installing a third-party skill collection is to keep equivalent instructions as project files in the repository you already version. The difference is where the text lives and who is accountable for it. Instructions inside your repository travel with the codebase and change in the same commit as the code they govern. An installed collection comes from tryproduck/produck-skills main, applies to every project you open with that agent, and moves when the vendor pushes, without anyone on your team reading the diff.

The trade runs the other way in one respect worth weighing: skills/taste/references/taste-principles.md is a rubric somebody else maintains, which means it arrives with considered revisions you would otherwise have to argue for yourself. Ownership cuts both ways. Note too that this collection is deliberately agent-agnostic, claiming 70+ agents and pointing Cursor, opencode and manual configurations at skills/produck-feedback/references/connect-mcp.md, so the choice is less about locking into one assistant and more about who reviews the prompt text your model reads on every run.

Editorial conclusion

Adopt the markdown half if you want your agent to write PRDs with bounded phases and acceptance criteria, or to run a screenshot pass over a landing page. Skip it if your feedback lives outside Produck, because user-alignment and taste stand alone while produck-feedback cannot reach any ticket store but the vendor's. Before relying on taste, read skills/taste/references/taste-principles.md and weigh it against a rubric your team already owns, since the benchmark table here is the authors grading their own skill on one fictional page. If demo-studio matters to you, confirm your agent can reach the HyperFrames renderer first, because it is not vendored in this repository and the film it renders shows an invented interface rather than your running product.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. Project website
4. README
5. tryproduck/produck-skills on GitHub