whychk — Claude Code skill: what it does, how to install
whychk — a skill for Claude Code and Codex that challenges freshly made choices. An isolated sub-session identifies unstated assumptions, demanding explanations one gap at a time.
MIT 123 installs ★ 1,470 updated 2026-10-25
npx skills add LilMGenius/paperthin --skill whychk -a claude-codeWhat whychk does
The skill runs solely upon explicit invocation right after a choice is made. The host restates the choice neutrally and passes it to a completely fresh, context-free sub-session without chat history or prior arguments. This outside critic determines what unstated premise must hold true for the decision to work.
Gaps are delivered strictly one per turn, ranked by impact. The user must provide their own reasoning to close each gap: the skill forbids providing answers, hints, or accepting deflections like "the model suggested it". When an explanation is shallow, the check narrows in on the vague part without expanding back to generic questions.
If the host environment cannot spawn a separate sub-session, the process halts immediately. Any unanswered gap turns into a persistent re-review flag written directly into the plan, ticket, or commit message.
What it helps with
- surface hidden assumptions behind an architectural choice
- verify that choices are grounded in real arguments
- flag unverified technical assumptions in commit messages
Example prompts for the agent
- Run whychk on this database choice
- whychk: review our decision to pick Fastify
- whychk for the auth migration decision
How to install whychk
1Skills CLI (Vercel)
npx skills add LilMGenius/paperthin --skill whychk -a claude-code2Manually: put the files in place
~/.claude/skills/whychk/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/whychk/
1Skills CLI (Vercel)
npx skills add LilMGenius/paperthin --skill whychk -a codex2Manually: put the files in place
.agents/skills/whychk/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add LilMGenius/paperthin --skill whychk -a cursor2Manually: put the files in place
.cursor/skills/whychk/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install LilMGenius/paperthin skills/depth/whychk2Manually: put the files in place
.github/skills/whychk/SKILL.md1Skills CLI (Vercel)
npx skills add LilMGenius/paperthin --skill whychk -a gemini-cli2Platform CLI
gemini skills install https://github.com/LilMGenius/paperthin --path skills/depth/whychkManually: put the files in place
.devin/skills/whychk/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add LilMGenius/paperthin --skill whychk -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about whychk
How to install whychk in Claude Code?
Run npx skills add LilMGenius/paperthin --skill whychk -a claude-code or place the definition at ~/.claude/skills/whychk/SKILL.md.
What does whychk do?
It feeds a chosen option to an isolated sub-session, spots the most fragile unstated assumption, and requires the user to defend the choice one gap at a time.
What is required for whychk to run?
The host environment must support launching independent sub-sessions without shared history; otherwise the skill stops.
Source text
SKILL.md · MIT
---
name: whychk
description: "Right after you pick an option, have a fresh outsider find the sharpest gap in the choice, then explain that gap in your own words, one gap at a time."
disable-model-invocation: true
---
Make the person who chose explain why, one gap at a time, before the decision counts as earned.
## Goal
A decision made after options were laid out without arguments can feel settled while resting on assumptions nobody said aloud. This check surfaces them one by one and lets the user close each with their own reasoning. It runs only when the user reaches for it, because a reflex that challenges every decision would breed chronic self-doubt.
## Workflow
1. Run when the user invokes this skill right after a decision, typically one picked from options presented without arguments; it never starts on its own. Restate the decision neutrally in one or two sentences, with no reasoning attached, and do not let the user borrow reasoning that was never given.
2. Launch a fresh, context-free sub-session that never watched the decision being made. Give it only the restated decision, never the options discussion, your reasoning, or the session history. Never use the session that made the choice as the critic; if the host cannot start a separate session, say so and stop rather than critique the choice yourself.
3. Ask the critic what a skeptical outsider would question, what is assumed but unstated, and what must be true for the choice to hold. Its answers are candidate gaps; rank them by how much of the decision falls if each one is wrong.
4. Take the sharpest unresolved gap from its answer and put exactly that one gap to the user, asking them to explain it in their own words. Never hand over a checklist.
5. Judge the answer. The gap closes only when the explanation resolves the actual question. If it restates the question, deflects, or leans on authority ("the agent suggested it"), it stays open.
6. After a shallow answer, narrow to the exact part that stayed vague and ask about that. Never repeat the same question and never widen back to a generic "are you sure?".
7. Continue with the next sharpest gap. Stop when every gap is explained, or when narrowing on one gap yields nothing new and the user leaves it open. Both endings are valid.
8. For an unexplained gap, name its specific assumption and flag the decision for re-review, written where the decision is recorded (its plan, issue or commit message) so the flag outlives the conversation. The flag means the decision is not yet earned, not that it was wrong. Never force a resolution and never let the user paper over the gap.
## Rules
- The critic is always a separate, context-free sub-session that receives the decision alone.
- Exactly one gap per turn, the sharpest unresolved one.
- Every reason that closes a gap comes from the user. This skill supplies none, not even as a hint or a leading question.
- Reject answers that restate the question, deflect, or cite authority; say why in one line.
- An open gap carries forward as a flag and is never dropped or softened.
- Stay neutral about whether the decision is right; the check measures understanding, not outcome.
## Verification
Before finishing:
1. The signal is the change between the user's first answer and the answer after narrowing, not the absence of objections.
2. Every closed gap traces to an explanation from the user that resolved the actual question.
3. Every open gap is listed with its specific assumption and carries forward as a flag for re-review.
4. No reason that closed a gap came from this skill.
© LilMGenius · MIT · LilMGenius/paperthin