bozhoudev/finding-unknowns-skills — explained in plain English
Analysis updated 2026-05-18
Add global behavioral rules to Claude Code so it asks better clarifying questions.
Reduce the chance of a long coding task going off track due to unstated assumptions.
Standardize a pre-implementation planning and post-implementation review workflow for a team.
Get a written implementation plan and deviation log for every non-trivial change.
| bozhoudev/finding-unknowns-skills | bozhoudev/xhs-article-to-images | core-trading/world-cup-trading-bot-ts | |
|---|---|---|---|
| Stars | 57 | 57 | 57 |
| Language | — | CSS | TypeScript |
| Setup difficulty | easy | moderate | moderate |
| Complexity | 1/5 | 2/5 | 3/5 |
| Audience | developer | writer | developer |
Figures from each repo's GitHub metadata at analysis time.
This repository is a single configuration file, not a piece of software you run, meant to change how Claude Code behaves when working on your codebase. It packages up a set of ideas from an article by an Anthropic team member about a recurring problem: when someone gives an AI coding assistant instructions, there are always things the person did not say, either because they forgot, did not know how to phrase it, or assumed it was obvious. The article calls these gaps unknowns, and argues that an assistant left to guess through them silently is how a long task quietly goes off track. The guidelines sort these gaps into four types: things clearly stated, things the person knows they have not figured out yet, things so obvious the person would never think to write them down but would recognize instantly if suggested, and things nobody involved has even considered. The file instructs Claude to actively surface the last three kinds instead of quietly guessing its way through them, using a set of named techniques at different points in a task, such as asking clarifying questions before starting, brainstorming and building small prototypes when a direction is unclear, writing an implementation plan before touching code, logging any deviations from that plan while working, and producing a short explainer plus quiz afterward so the person can confirm what actually changed. Installing it means appending the contents of the file to a Claude Code configuration file, either globally so it applies to every project you work on, or to a single project's own configuration file, using a one line download command. There is no code to compile or dependencies to install, since the entire tool is instructional text read by the AI assistant itself. The author is explicit that this approach favors thoroughness over speed, and that Claude should skip the extra ceremony for trivial changes like fixing a typo. This project is aimed at developers who use Claude Code regularly and want more reliable results on larger, more ambiguous tasks, and requires no technical setup beyond copying text into a configuration file.
A drop-in Claude Code configuration file that trains the AI assistant to actively surface unstated assumptions before and during a task instead of silently guessing.
Licensed under MIT, which allows free use, modification, and commercial use as long as you keep the copyright notice.
Setup difficulty is rated easy, with roughly 5min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
double-check against the repo, no cap.