explain · insilico-org — Claude Code skill: what it does, how to install
A Claude Code skill for turning research findings into competing, testable explanations. It activates when plsi research next selects explain.
MIT ★ 93 updated 2026-11-15
npx skills add Insilico-org/plsi --skill explain -a claude-codeWhat explain does
The skill turns research findings into competing hypotheses and records them in a graph. It considers chance or no effect, confounding, and measurement artifacts before moving to mechanistic explanations. Each surviving hypothesis includes an observable result that would show it is wrong.
It writes hypothesis nodes with a question, an if_false condition, and a tested_by link when the relevant question already exists. Set-aside candidates go in dropped with a reason; the skill does not issue verdicts. It needs relevant findings from next; when the search itself was wrong, the instruction is to gather instead. Node structure comes from shared/research-graph, while persona.md and shared/team provide working norms.
What it helps with
- turn research findings into testable explanations
- distinguish mechanisms from chance and measurement artifacts
- state what observation would falsify a hypothesis
- record discarded explanations and why they were set aside
Example prompts for the agent
- Turn the findings from
plsi research nextinto competing explanations. - For each hypothesis, name an observation that would show it is wrong.
- Separate testable hypotheses from candidates that should be set aside.
How to install explain
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill explain -a claude-code2Manually: put the files in place
~/.claude/skills/explain/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/explain/
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill explain -a codex2Manually: put the files in place
.agents/skills/explain/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill explain -a cursor2Manually: put the files in place
.cursor/skills/explain/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install Insilico-org/plsi skills/lab/explain2Manually: put the files in place
.github/skills/explain/SKILL.md1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill explain -a gemini-cli2Platform CLI
gemini skills install https://github.com/Insilico-org/plsi --path skills/lab/explainManually: put the files in place
.devin/skills/explain/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.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 explain
How do I install explain in Claude Code?
Run npx skills add Insilico-org/plsi --skill explain -a claude-code. The skill file is at ~/.claude/skills/explain/SKILL.md.
What does the explain skill do?
It builds competing hypotheses from research findings, describes an observable way each could be disproved, and saves the candidates in the research graph.
How does explain handle hypotheses that cannot be distinguished?
It compares candidates in pairs. If no observation could tell two explanations apart, they are merged; each surviving hypothesis gets an observable falsifier.
Source text
SKILL.md · MIT
--- name: explain description: Use as PLSI Research Orchestrator when `plsi research next` says the move is `explain` - put the competing accounts of the evidence on the table, the boring ones first, each with what we would see if it were false. --- # Explain Follow persona.md norms and `shared/team`. Owner: **PLSI Research Orchestrator**. Writes `hypothesis` nodes; the shape is in `shared/research-graph`. `next` has handed you the question, every finding that bears on it, and each one's `note`. Nothing here needs a search. Explain is due only where findings exist. Where they say nothing useful, explaining the absence is legitimate (a tooling limit, a question nobody asked, a result only in tables this loop cannot read), and each is a falsifiable candidate. If the honest answer is that the search was wrong, do a gather instead and save that. When the candidates feel thin, or you cannot tell whether a falsifier is any good, read `lab/good-hypothesis`. ## 1. The boring three first, as a checklist Before any mechanism, consider: 1. **Nothing**: chance, or no real effect. 2. **Confounding**: something else accounts for it. 3. **Artifact**: measurement, selection or reporting produced it. Then the mechanistic candidates. Write the boring ones that are live as nodes and name the ones you dismissed in `dropped`, with why. ## 2. Every candidate that survives the separating test, each with `if_false` There is no fixed number. **Two candidates that no observation could separate are one** (`lab/test` §3): run that test over every pair, merge what it catches, and keep what survives. A question that needs many more is split in Ask, not crowded here. `next` hands back `dropped` in `already_tried`, so a set-aside candidate stays set aside. **Every candidate records what we would see if it were wrong** (refusal 2). An `if_false` must name something observable: *"the symptom began before the prescription date in the user's own log"*, not *"the hypothesis would not hold"*. Fill `tested_by` only with a question that **already exists** (refusal 6 rejects a link to a missing node). Otherwise leave it empty and say in the body what would settle it; Test fills the link when it creates that question. Never point `tested_by` back at the question that proposed the hypothesis. ## 3. Sharpening, not nesting Hypotheses do not nest. A candidate sharpened after new evidence is a **new** hypothesis under the same question, with the old one closed `superseded` and cited. Structure lives in the question tree. ## 4. Read your way to hypotheses A topic holds tens of findings, far too few for graph-structure hypothesis generation; candidates come from reading the findings. The graph stops you asking the same question twice and counts what you find. ## 5. Save `nodes` of `kind: hypothesis`, each with its `question`, `if_false`, `tested_by` where you know it, and a body of one or two lines saying what it claims. In `dropped`, the candidates you set aside, with why. ## Rules - You never write a verdict, and never a verdict word as a status (refusal 7). - Do not propose a candidate nothing could test. Put it in `open` as a direction that needs a measurement nobody has.
© Insilico-org · MIT · Insilico-org/plsi
Similar skills
All skills for Science and research →explain by other authors comes first.