Detailed Analysis
A Reddit user has released murPick, an open-source "agent skill" designed to address a subtle but consequential behavior in Claude Code: the tendency for the model to silently collapse an open-ended research task into a fixed set of product decisions. According to the developer's baseline testing, when Claude Code was asked to research a product space without any intervening skill, it would produce a thorough report but then unilaterally conclude with prescriptive recommendations — in one documented case, dictating "V1 = one skeleton + three modules" alongside a definitive list of features not to build. The user never selected these features themselves; Claude simply decided on their behalf, burying the decision-making inside a wall of research prose rather than surfacing it as a choice.
murPick's fix is structurally simple: rather than letting Claude synthesize research directly into a recommendation, the skill converts the research output into an interactive, self-contained HTML file that functions as an à la carte feature menu. Every candidate feature identified during research becomes a checkbox. The file works on mobile, auto-saves state, and includes a "Save checked copy" button that embeds the user's selections directly into the file, which can then be fed back to Claude for a follow-up "selection report." Critically, features the user does not check aren't simply forgotten — they become an explicit "not building this" list, preserving the research work while making the omission a visible, intentional artifact rather than something quietly dropped in a paragraph nobody re-reads.
This small tool speaks to a broader and increasingly discussed friction point in agentic coding workflows: as AI coding assistants like Claude Code become more capable of autonomous, multi-step reasoning, they also become more capable of making high-leverage decisions without adequate checkpoints for human input. Research and planning are exactly the phases where a developer wants maximum optionality and minimal premature commitment, yet LLMs — optimized to produce confident, coherent, and complete-feeling outputs — have a structural bias toward resolving ambiguity themselves rather than leaving it open for the user. murPick is a lightweight, non-intrusive attempt to reintroduce a human decision gate into that pipeline, without requiring the user to abandon the efficiency of having Claude do the research legwork in the first place.
The project is also notable for how it was built and validated: the developer describes documenting baseline failures verbatim, "pressure-testing" the skill, and gating every generated menu behind a 12-check validator, suggesting an engineering-rigor approach to what could otherwise be a purely cosmetic UI wrapper. This reflects a maturing pattern within the Claude Code ecosystem, where third-party "skills" — modular extensions to Claude's agentic behavior — are increasingly built with the same testing discipline as production software, rather than as one-off prompts. It also reflects growing community sophistication in identifying and correcting specific failure modes of agentic AI tools, rather than treating them as black boxes to be used as-is. The self-reported limitation — that menus with 100+ items become unwieldy on mobile — signals that this is an early-stage, community-tested tool rather than a polished product, and underscores that the core challenge of presenting large decision spaces to humans without overwhelming them remains only partially solved, even as tools like murPick chip away at Claude's tendency to make decisions users never explicitly delegated to it.
Read original article →