murPick

AI research, à la carte.

Don't read the report — order from it. You pick the dishes, AI does the cooking, and nothing off-menu makes it to the table.

Ask an AI to research competing products and you get a 5,000-word report. By the time you reach the end you've forgotten the beginning — and somewhere in the middle, the AI quietly decided your feature set for you.

murPick is an Agent Skill that turns product research into something you already know how to use: a menu. The AI digests your reference materials, researches the field, then serves every candidate feature as a checkbox. You order like you're at a restaurant. What you check goes into the spec. What you don't check becomes explicit boundaries — "decided against" instead of "forgotten".

How it works

1. Digest — you drop in screenshots, screen recordings, links, or a folder. The skill inspects everything (video frames included), renames files descriptively, dedupes, and identifies each product.
2. Research — open-source counterparts, commercial leaders, category overview. With honesty rules: features that demo well but demand manual data entry get flagged, real-world API/ToS barriers get named.
3. The menu — a single self-contained HTML file. Every feature is a checkbox with a stable id, grouped into product cards with source links. Auto-saves as you tap (localStorage). Works on your phone.
4. The report — send the checked file back. Checked items map into capability domains — none dropped. Unchecked items become the boundary list. You get an HTML report with decision tables, a pure-CSS architecture diagram, and an honest risk table.

The round trip is the trick: the menu's "Save as checked copy" button burns your checkbox state into the HTML itself (data-burned attributes) and downloads it. No account, no server, no copy-pasting a wall of Markdown from your phone. The file is the data.

Why a menu?

Because the failure mode of AI-assisted product research isn't bad research — it's decision laundering. In our baseline tests (same materials, same prompts, no skill), the agent produced a competent report that ended with:

"V1 = one skeleton + three modules + one principle… Definitive don't-build list (write it into the README to prevent scope creep)"

The user never picked anything. The AI picked. With murPick, the same agent produced a 97-item menu and handed the choosing back to the human. Granularity matters too: the no-skill baseline compressed the field into 19 theme-level bullets; the menu preserved 97–143 individually checkable features.

Install

Claude Code:

git clone https://github.com/ymustc/murPick.git ~/.claude/skills/murpick

Runtimes that read ~/.agents/skills/ (Codex, Copilot CLI, Gemini CLI):

git clone https://github.com/ymustc/murPick.git ~/.agents/skills/murpick

Compatibility note: murPick follows the open Agent Skills spec and its pipeline only needs common tools (shell, a web-fetch tool, ffmpeg for videos, Node for the validator). It is battle-tested on Claude Code; on other runtimes it should work but hasn't been formally tested — reports and issues are very welcome. It also works with any Anthropic-compatible model endpoint behind Claude Code.

Use

I collected screenshots of habit-tracker apps I like in ./refs — help me research this space, I want to decide what my own app's v1 should include.

or simply:

Make me a feature menu from this research.

When you've checked your picks, hit Save as checked copy and send the downloaded file back:

Here's my checked menu — write the selection report.

No-install fallback

A useful chunk of the discipline works as a plain prompt convention: split research and selection into two passes, ban the research pass from recommending anything, and require the rejects to be listed as explicit boundaries (credit: launch-thread feedback). What the skill's artifact adds on top is per-item granularity, the burn-and-save round trip, and a validator-checked guarantee that every unchecked item resurfaces in the report.

Quality: tested like code, not vibed

This skill was built with TDD for documentation (RED-GREEN-REFACTOR): baseline runs without the skill documented real failures verbatim; every clause targets one; pressure tests (time pressure, ill-fitting materials) closed the loopholes.

Machine-checkable requirements are enforced by script, not prose: scripts/validate-menu.js runs 14 hard checks on every generated menu (checkbox id uniqueness, burn-and-save logic, localStorage, self-containment…). In our test suite it fails the no-skill baseline menu 7/14 and passes every skill-generated menu including a real 143-item case.

Examples

See examples/ for a generated menu, a checked copy, and the resulting report.

---

Part of the mur toolkit by Miao YU. 中文说明见 README.zh-CN.md。

MIT License.