Cookiy AI user-research-skill: an agent skill for interviews and surveys

What the Cookiy AI skill actually does for an agent

The repository is a skill for AI agents, not a standalone research application. According to the README, it lets an agent plan, run and synthesize user research without leaving the conversation, covering qualitative interviews and quantitative surveys. The README splits this into four capabilities: research planning (plans, screening questionnaires, interview guides), report synthesis (turning raw transcripts into reports with codebooks, personas and prioritized findings), qualitative studies through AI-moderated interviews, and quantitative surveys with conditional logic and respondent recruitment.

The audience is narrow and specific. It is people who already work inside an agent such as Claude Code, Codex, Cursor or OpenClaw and want research artifacts produced as part of that session. The README positions the repository as the skill, while the full product lives at cookiy.ai, and notes that once you log in on the website the web app and the CLI skill stay in sync across studies, participants and results. So the skill is a front end to a hosted service for the parts that involve real participants.

Where the work runs: local synthesis versus hosted studies

The architecture implied by the repository has two halves. The first is local to your agent: planning documents and transcript synthesis. The SKILL.md file at the repository root and the references directory are what the agent reads to decide how to behave, which is the standard layout for an agent skill. Nothing in the README suggests those steps require a network call, and the examples treat synthesis as a self-contained task: hand the agent twenty transcripts and it runs a five-phase synthesis pipeline producing themes and personas.

The second half is hosted. Qualitative studies with real or synthetic participants and quantitative surveys both go through Cookiy AI, and the README ties the web app and the CLI skill to the same account so studies, participants and results stay in sync. The network configuration in the install steps points at a single host, s-api.cookiy.ai, which is the only endpoint named anywhere in the README. That is the boundary to keep in mind: the skill is open source and MIT licensed, but the participant-facing machinery is a service, and the README does not describe a self-hosted alternative for it.

Installing the skill in Claude Code, Codex or Cursor

Claude Code installs it as a plugin from a marketplace. The README gives three commands: add the marketplace, install the plugin, then reload. The plugin name in the install command is user-research@cookiy-ai, which differs from the repository name, so copy it exactly.

/plugin marketplace add cookiy-ai/user-research-skill
/plugin install user-research@cookiy-ai
/reload-plugins

After the reload, the README says you can either type /user-research-cookiy to invoke the skill explicitly or simply describe your research goal and let the agent load the skill by semantic matching. There is an optional auto-update path: run /plugin, open the Marketplaces tab, select cookiy-ai and enable auto-update. For Codex, Cursor, OpenClaw and other agents the README gives a single command, npx cookiy-ai, or you follow your agent's own skill installation instructions.

npx cookiy-ai

For the Claude desktop applications the flow is graphical. In Claude Chat Desktop you download the skill ZIP from the latest release, then use Customize, Skills, Create Skill, Upload a skill. In Claude Cowork or Code Desktop you add the marketplace by entering cookiy-ai/user-research-skill, sync, then enable the plugin under Plugins, Personal. Both desktop paths require turning on network egress under Settings, Capabilities, Code execution and file creation, and the README says to either set the domain allowlist to All domains or add s-api.cookiy.ai under additional allowed domains. That last detail matters: if you keep the allowlist narrow, add the API host or the hosted study features will not reach the service.

A first real use: from a churn question to a research plan

The README's own example is the cleanest first task. Describe the goal and let the agent pick up the skill.

I want to understand why users churn after onboarding

According to the README, this produces a research plan, a screener and an interview guide aimed at early-stage drop-off. Nothing leaves your machine for that step, and it is the right way to judge whether the skill's output style suits you before connecting an account.

The second example is synthesis, which is the other half that runs locally.

Here are 20 interview transcripts, give me a report

The README states this runs a five-phase synthesis pipeline and returns a structured report with themes and personas. The repository does not document the phases by name, so if you need to know what happens between transcript and report, read SKILL.md and the references directory before relying on the output.

The hosted path looks the same from the prompt side but behaves differently underneath.

Run a study on how developers choose CI/CD tools

Per the README, this creates the study through Cookiy AI, runs AI-moderated interviews and delivers the report. The survey example is "Survey 200 users about feature satisfaction", which designs the survey, recruits respondents and collects results. Both require the network configuration described above and an account on cookiy.ai.

Limitations: hosted participants, thin documentation, and the wrong use cases

The clearest limitation is the dependency. Anything involving real or synthetic participants, recruitment or survey fielding goes through Cookiy AI. The repository is MIT licensed and you can read every file in it, but the README does not describe running those studies against your own infrastructure, and the only named endpoint is s-api.cookiy.ai. If your organization cannot send participant data to a third-party service, the qualitative and quantitative halves are simply unavailable to you, and what remains is planning and transcript synthesis.

Documentation depth is the second issue. The README is strong on installation and examples and thin on mechanics. It names a five-phase synthesis pipeline without listing the phases. It does not document rollback, rate limits, data retention, or what happens to a study if the network allowlist blocks the API mid-run. The repository does contain SKILL.md and a references directory, which is where the actual instructions to the agent live, so the README is not the whole story, but a reader who only reads the README will not know how the synthesis stages are structured or how to audit them.

The third issue is scope. This is a tool for producing research artifacts and, through the service, for collecting responses. It is not a repository, a consent management system, or a substitute for a researcher deciding what to ask. The README's framing of the project as the human layer in the AI stack is a positioning statement, not a description of validation. If your study needs regulatory sign-off, recorded consent trails or a defensible sampling frame, the README gives you nothing to work with.

How it compares with a general-purpose agent workflow

The realistic alternative is not another research skill. It is using your agent with no skill installed: you write the interview guide yourself, paste transcripts into the chat, and ask for a summary. The difference is that a skill ships a fixed procedure. The agent reads SKILL.md and the references directory and follows a defined pipeline instead of improvising a prompt each time, which is what makes the output consistent across sessions and across team members. The cost of that consistency is that you inherit someone else's procedure, including its five-phase synthesis structure, and you cannot tune it without editing the skill.

The second alternative is the Cookiy web app on its own, without the agent. The README treats these as connected rather than competing: log in on the website and the web app and the CLI skill stay in sync across studies, participants and results. So the real choice is where you want to sit. If you prefer a graphical interface for building studies and reviewing participants, the website covers that and the skill adds nothing. If you work in a terminal and want plans and synthesis to appear where the rest of your engineering work happens, the skill is the entry point and the website becomes the place you go for participant management.

Licence, maintenance and what an upgrade costs you

The repository is MIT licensed, Copyright Cookiy AI, and the LICENSE file sits at the top level alongside README.md, SKILL.md and the plugins directory. MIT is permissive: you can copy, modify and redistribute the skill files, including inside a commercial product, provided the copyright notice and permission notice travel with them. That covers the repository contents. It does not cover the hosted service, and the README does not state terms for the API, so treat the licence as governing the skill code only. This is not legal advice; if you plan to redistribute the skill inside a product, have someone read the LICENSE file and the terms for the service separately.

On maintenance, the last push to the default branch was on 2026-08-19, and the repository is not archived. The most recent release, tagged latest and offering a downloadable skill ZIP, dates from 2026-04-16. Those two dates tell you the code has moved since the last packaged release, which is normal for a skill repository but worth knowing if you install from the ZIP rather than from the marketplace or npx. Upgrade cost is low by design: the Claude Code path has an auto-update toggle under the Marketplaces tab, and the npx path fetches the current version when you run it. The desktop upload path is the manual one, and there the gap between the April release and the August push is the thing to check before you assume the ZIP matches the repository.

Editorial conclusion

Adopt it if your research work already happens inside Claude Code, Codex, Cursor or OpenClaw and you want plans, guides and synthesis produced in the same session, and if you accept that live interviews and recruitment route through Cookiy AI servers. Do not adopt it if you need a fully local or self-hosted pipeline, or if you cannot send participant data to a third party. Before committing, verify the network egress settings for your client, confirm which parts of the workflow still function without a Cookiy AI account, and read SKILL.md and the references directory in the repository to see what the skill actually instructs the agent to do.

Frequently asked questions

Official sources

1. cookiy-ai/user-research-skill on GitHub
2. License: MIT
3. Project website
4. README
5. Releases