Guide · Agent skill
How to use the Product Direction Judge agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The Product Direction Judge is an agent skill that checks feature requests, feature pull requests and your own feature proposals against a repository's written product direction. It returns a verdict with a quoted line as evidence and leaves the final decision to a maintainer. This guide explains what the skill does, how to install it, how to trigger it with ordinary requests, and what to watch for, so you can use it on your own project.
Product Direction Judgeprisma/orm · Apache-2.0 Download the skill Quick start
- 1Download the ZIP. Get the product-direction ZIP from the skill's page on this directory.
- 2Install the skill. Unzip it into ~/.claude/skills/ for personal use, or into the project's .claude/skills/ folder, or upload it under Settings → Capabilities → Skills in the Claude apps.
- 3Confirm the repository has direction pages. Check for an agent instructions file, project intent page, vision page, decision records and a steering list. Missing parts are reported as unknown.
- 4Make a natural request. Ask the agent to judge a feature request, triage a pull request, or check a proposal against the direction docs.
- 5Read the cited verdict. Check the verdict, the reason and the linked quote before acting on it.
- 6Leave the decision to a maintainer. Use the output as input. A maintainer makes the final call, and any decline needs an accepted record or a maintainer's own linked words.
What does the product-direction skill make the agent do?
Before reviewing code, the agent reads the project's written intent, its vision, and its decision records. It then gives one verdict for the request: aligned, rejected, unclear or not_applicable. Each verdict comes with a one-sentence reason and an anchor, which is a link and a quote from the repository's own pages.
The skill is built to be conservative. If the written record does not clearly speak to a request, the verdict is unclear. Bugs and routine code review are marked not_applicable, because a judgment about direction means nothing there.
- aligned: the work advances a stated value or durable theme.
- rejected: it conflicts with a value or an accepted decision, or it belongs elsewhere, such as an extension or user code.
- unclear: the record does not clearly speak to it. Often the verdict names the question a maintainer must answer.
- not_applicable: a bug, or anything where direction does not apply.
Install the skill in Claude
Download the ZIP from the skill's page on this directory. Then choose one of two installation routes, depending on whether you want the skill for yourself or for one project.
For personal use, unzip the download into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder. You can also upload the ZIP in the Claude apps under Settings → Capabilities → Skills.
- Personal install: unzip into ~/.claude/skills/.
- Project install: unzip into the project's .claude/skills/ folder.
- Claude apps: upload the ZIP under Settings → Capabilities → Skills.
Check that the repository has a direction record
The skill looks for four parts of a written record in the repository, plus its own reading rules. The default paths are docs/direction/project-intent.md, docs/direction/vision.md, and docs/direction/decisions/ with an index at docs/direction/decisions/index.md. The agent first reads the repository's agent instructions file, such as AGENTS.md or CLAUDE.md, and follows any pointer it finds there.
If the repository already keeps architecture decision records, the skill uses that house format and numbering rather than creating a new layout. The default paths are used only when the agent creates a part that does not exist.
- Project intent: what the project is, who it is for, and what it does not need to optimise for.
- Vision: values and durable themes, with nothing about current work.
- Decision records: one per no, with what to use instead and what would reopen it.
- Steering list: accepted nos, one line each, read before proposing anything.
Trigger the skill with natural requests
You do not need to name the skill. The description covers triage, questions about project direction, maintainer decisions, and proposals. Phrase your request the way you would to a colleague, and say what the agent should look at.
The skill also applies before the agent proposes any feature, API, command, config option, dependency or abstraction, even when nobody mentions direction. That is useful when an agent is drafting code for you.
- "Does this feature request fit the project? Check it against the direction docs."
- "Triage the open feature pull requests and give each a verdict with a cited line."
- "Before you add a config option for this, check whether the project already wants it."
- "Set up a project intent page and a steering list for this repository."
- "What does this project say it will not build?"
Read a verdict and what to do next
A verdict is a short structured block. The agent fills in the fields visionFit, visionFitReason, visionFitEvidence and bestSolution. If a shipped feature already covers the central ask, bestSolution begins with "Already exists:" and names that feature.
The skill does not decline requests on its own. A decline needs an accepted decision record that covers the request, or a maintainer's own earlier words on the issue, linked. Proposed declines state a confidence level, and anything below high confidence is given as unclear. Treat the output as a draft for a maintainer to review.
- Read the visionFitEvidence link before acting on the verdict.
- A rejected verdict without grounds for a decline leaves the request open, with a note on where the work belongs.
- A Proposed decision record never grounds a verdict. A maintainer must accept or delete it.
Set up or update the direction pages
The skill also helps create or update the pages that hold direction. Ask the agent to set up a project intent page, a vision page, decision records, or a steering list. It uses the templates included in the skill's files, and it keeps the repository's existing layout where one exists.
A decision record is written after a maintainer has decided not to build something that would otherwise be proposed again. Records are kept short, and the Context section quotes the deciding maintainer with the date.
- project-intent.md: the project's identity, audience and what not to assume.
- vision.md: values and themes, linking to the skill rather than copying its rules.
- decisions/: one record per accepted or proposed no, using the decision-record template.
- decisions/index.md: the steering list of accepted nos with what to use instead.
What people use it for
- Triaging a backlog of feature requests and labelling each with a cited verdict.
- Checking a feature idea against the project's direction before writing any code.
- Stopping an agent from adding a config option or dependency the project has already declined.
- Setting up a project intent page, vision page, decision records and a steering list for a team.
- Writing a decision record that quotes a maintainer's earlier reasoning after a no.
Frequently asked questions
How do I install the product-direction skill?
Download the ZIP from the skill's page on this directory. For personal use, unzip it into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder, or upload the ZIP under Settings → Capabilities → Skills in the Claude apps.
What does the skill require to work?
It needs access to the repository files, since it reads SKILL.md and its reference and template markdown files. It also relies on the repository's issue and pull request history to find earlier maintainer comments, although the files do not specify this. It runs no scripts and uses no network access, according to the source.
Is the product-direction skill free to use?
The source does not state pricing or licence terms for the skill's use beyond its licence, which is Apache-2.0. Check the licence text for the terms that apply.
Why does the skill return unclear so often?
The skill is designed to be conservative. If the repository has no written direction pages, or the pages do not clearly speak to the request, it reports the missing parts as unknown and returns unclear rather than inventing direction.
Can the skill reject feature requests automatically?
No. The skill says a wrong no drives people away, so a decline needs an accepted decision record or a maintainer's own linked words. A maintainer makes the final decision.
What is the difference between this skill and reviewing code?
The skill decides whether a request fits the project's direction, before a detailed code review. Bugs and routine code review are marked not_applicable, since a direction judgment means nothing for them.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
This skill suits maintainers who want consistent, cited answers when triaging feature requests, and contributors who want to check an idea against written direction before writing code. It is a poor fit for bug triage, routine code review, or repositories with no written direction pages, where most verdicts will be unclear. Use it as a drafting aid for a maintainer, not as an automatic rejection tool.
Download the skill