Produck Skills: agent skills that turn messy user requests into executable PRDs

The gap Produck Skills targets: agents that build the wrong thing quickly

Coding agents are good at executing a clear instruction and bad at recovering from a vague one. A request like "make onboarding better" gives the agent nothing to test against, so it produces plausible code with no acceptance criteria. Produck Skills is a set of markdown folders that teach an agent how to convert that kind of request into a PRD with scope, bounded phases, testable acceptance criteria and explicit do-not-do boundaries.

The audience is narrow by design. The README describes the collection as documentation-driven skills for coding agents, and lists Claude Code, Cursor, Codex and Copilot as targets. If you are a solo developer writing your own tickets, the user-alignment skill duplicates work you already do in your head. It pays off when several people and an agent share the same brief and disagree about what "done" means. The repository also ships a taste skill for visual work and a demo-studio skill for launch videos, so the collection is broader than PRD writing, but user-alignment is the one the README leads with.

How the skills are structured and what actually runs

There is no runtime. Each skill is a folder of markdown under skills/, with a SKILL.md file at its root and optional references/ files beside it. The agent reads those files as context when the skill is invoked. That is the whole mechanism: the repository changes what the model knows at the moment it plans, not what it can execute.

The one exception is produck-feedback, which drives the Produck MCP, a hosted server exposing two tools. search_feedback finds relevant feedback tickets and can filter by domain. get_feedback loads a single ticket's full context, which the README describes as the user's written or spoken feedback, reconstructed page HTML, annotation markers, session activity and a generated design doc. The skill then drives a feedback to aligned PRD to build loop and pairs with user-alignment. Everything else in the catalog is static text.

That split matters for evaluation. The markdown skills either fit your workflow or they do not, and you can read them before installing. The MCP-backed skill is a different proposition: its usefulness depends on feedback already sitting in a Produck account.

Installing Produck Skills and running user-alignment once

The README gives a single install line. It uses the skills CLI through npx, so nothing is installed globally, and the README states the skills are available the next time your agent starts.

npx skills add tryproduck/produck-skills

After that, the new folders should appear wherever your agent keeps skills, and a restart of the agent should make them selectable. The README does not document an uninstall or update command, so plan on removing folders by hand if you change your mind.

If you want the feedback loop rather than just the PRD skill, the Produck MCP needs a token first. The README says to mint one at tryproduck.com/app under MCP, and that it looks like mcp_live_... . For Claude Code the setup writes the token to a file:

mkdir -p ~/.produck && echo "mcp_live_YOUR_TOKEN" > ~/.produck/mcp_token

Then, inside Claude Code, the plugin marketplace is added and the plugin installed:

/plugin marketplace add tryproduck/produck-skills
/plugin install produck@produck

For Codex the token goes into an environment variable instead, and the marketplace is added from the shell:

export PRODUCK_MCP_TOKEN=mcp_live_YOUR_TOKEN
codex plugin marketplace add tryproduck/produck-skills

The README notes that the token must be set before the MCP tools will work, and that the skills themselves work without it. For Cursor, opencode or a manual Codex and Claude Code setup, the README points to skills/produck-feedback/references/connect-mcp.md for the snippets.

The taste benchmark is a demo, not a measurement

The README includes a before-and-after comparison of a fictional SaaS landing page, scored across nine dimensions: category idea, worldview, hierarchy, clarity, trust, rhythm, CTA gravity, domain specificity and consistency. The baseline scores 1 or 2 on every dimension; the taste-guided rebuild scores 4 or 5. Screenshots were captured at 1440 x 1100 with npx -y chrome-devtools-axi and combined into one asset.

Read that table as an illustration of what the taste skill asks for, not as evidence that it improves conversion. The page is fictional, the rubric is defined in skills/taste/references/taste-principles.md by the same team that built the skill, and the scoring is self-administered. The benchmark does show something concrete: the taste skill is opinionated about specifics such as domain artifacts (opportunity memos, accepted and rejected evidence, decision briefs) rather than generic polish advice. Whether those specifics fit your product is a judgement the table cannot make for you.

Where Produck Skills is the wrong tool

Three limits are visible in the repository itself.

First, the feedback path is tied to one vendor. produck-feedback reads tickets through the Produck MCP, so if your feedback lives in a support desk, a spreadsheet or a public issue tracker, the skill has nothing to read. The README does not describe an adapter or an export path for other sources.

Second, the token is a manual step. Both plugin installs require a token minted in the Produck app, and the README states the MCP tools will not work until it is set. That is a per-developer setup task, and the README does not document rotation, expiry or what happens when the token is wrong.

Third, the collection is small and the README says more skills are on the way. Four skills cover PRD alignment, feedback triage, demo video production and visual taste. If your problem is test coverage, migration planning or release notes, none of these folders addresses it. The repository is also documentation-driven, which means its value tracks the quality of the markdown rather than any code that can be tested. There are no releases listed, so pinning to a version is not an option; you take the default branch as it stands.

How it differs from plain prompt libraries and spec-driven toolkits

The closest comparison is a prompt or rules file you write yourself and commit next to your code. That approach gives you full control and no dependency, but it is one file that grows until nobody reads it, and it has no notion of separate skills with their own references. Produck Skills splits the material into folders with a documented authoring contract in CONTRIBUTING.md covering folder layout, required frontmatter and a PR checklist. The trade-off is that you inherit someone else's structure and the README's stated bar of high-quality, UX-centered skills only.

A second comparison is spec-driven development tooling, which typically asks you to write a specification in a fixed format and then generates tasks from it. Produck Skills works in the opposite direction: it takes a messy inbound request, whether typed by a human or pulled from a feedback ticket, and produces the PRD. The two are complementary rather than competing, and the README does not position the collection against any named tool.

Licence, maintenance and the cost of upgrading

The repository is licensed Apache-2.0, which permits commercial use and modification and includes an explicit patent grant. If you fork a skill and ship it inside a product, keep the licence and notice files with the redistributed material. That is a description of the licence text, not legal advice; check with your own counsel before redistributing.

The last push to the default branch was on 2026-08-14, and the repository is not archived. There are no retrieved releases, so there is no version to pin and no changelog to read before an update. Because the installed artifact is markdown that your agent reads, an update changes agent behaviour without changing any code you can review in a diff of your own project. The practical upgrade cost is re-reading the SKILL.md files you rely on before pulling, and keeping your own edits in a separate branch so they do not collide with upstream changes.

Editorial conclusion

Adopt Produck Skills if your agent already writes code faster than your team can write acceptance criteria, and you want a user-alignment skill that produces bounded, testable PRDs rather than another prompt snippet. Skip it if you need a standalone feedback store: the produck-feedback skill reads tickets through the hosted Produck MCP and will not work without a token minted at tryproduck.com/app. Before rolling it out, check the frontmatter and folder layout in skills/user-alignment/SKILL.md against CONTRIBUTING.md, because the repository is documentation-driven and the install only copies what is in the skills/ directory.

Frequently asked questions

Official sources

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