ask · insilico-org — Claude Code research skill: questions and setup
This Claude Code skill turns a research direction into a question that search can settle, or breaks it into child questions. It activates when plsi research next selects ask.
MIT ★ 93 updated 2026-11-15
npx skills add Insilico-org/plsi --skill ask -a claude-codeWhat ask does
The skill writes testable research questions as question nodes. Before a search, it checks whether evidence could exist, whether either answer is possible, whether the subject and measurement are bounded, and whether the result would change the main conclusion.
A broad direction becomes three to five questions spanning at least three angles, such as mechanism, magnitude, or harm. A narrow direction gets one question with search terms. For a direction with a named subject, the skill can run plsi research discover --about "<the subject>" to distinguish well-studied relations from candidates. Saved nodes include stakes, angle, and parent links; a claim that a topic has not been studied requires last_search with two rounds using different vocabularies.
What it helps with
- turn a broad research direction into a searchable question
- split a topic across distinct research angles
- filter out questions whose answers are already known
- record how child questions relate to a parent
Example prompts for the agent
- Split this research direction into three to five questions from different angles.
- Write a focused question about how a gene affects the outcome, with search terms.
- Check which relations have been studied for the subject "<the subject>".
How to install ask
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill ask -a claude-code2Manually: put the files in place
~/.claude/skills/ask/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/ask/
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill ask -a codex2Manually: put the files in place
.agents/skills/ask/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill ask -a cursor2Manually: put the files in place
.cursor/skills/ask/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install Insilico-org/plsi skills/lab/ask2Manually: put the files in place
.github/skills/ask/SKILL.md1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill ask -a gemini-cli2Platform CLI
gemini skills install https://github.com/Insilico-org/plsi --path skills/lab/askManually: put the files in place
.devin/skills/ask/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 ask
How do I install ask in Claude Code?
Run npx skills add Insilico-org/plsi --skill ask -a claude-code. The skill file is at ~/.claude/skills/ask/SKILL.md.
What does ask do?
It turns a research direction into a question that search can resolve, or divides a broad topic into child questions. It saves those questions as nodes in the research graph.
When does ask split a direction into child questions?
When one research cycle cannot settle the direction, it creates three to five questions across at least three angles. For a narrow direction, it writes one question with search terms instead.
Source text
SKILL.md · MIT
--- name: ask description: Use as PLSI Research Orchestrator when `plsi research next` says the move is `ask` - turn the top frontier direction into one question a search could settle, or split it into three to five children across different angles. --- # Ask Follow persona.md norms and `shared/team`. Owner: **PLSI Research Orchestrator**. Writes `question` nodes; the shape is in `shared/research-graph`. A frontier item is a direction; a question is something a search can settle. Most of what the rest of the cycle costs is decided here. ## 1. Four tests, before any search A question that fails any of these is reworked, not searched. 1. **Could the evidence exist?** Something a study design could have measured. 2. **Can it come out both ways?** If one answer is already known, it is not a question. 3. **Is it bounded?** Who or what (population, system, sample), what was varied (exposure, marker, intervention), measured how (outcome). Without all three a finding cannot say where it came from. 4. **Does it change anything?** If either answer leaves the question in chief where it was, drop it. For test 2, when the direction has a subject (a drug, a gene, a condition), ask the field first: ```bash plsi research discover --about "<the subject>" ``` `well_studied` lists what the field has answered: do not write those children. `candidates` are relations standing on a handful of papers, each with PMIDs to aim Gather at. It is optional, and it proposes only: a relation unrelated to the question in chief is a different topic. `lab/deep-search` says what the counts mean. ## 2. Split it, or sharpen it **Split** when the direction is too big to settle in one cycle: three to five children, **from at least three different angles**. | Angle | Asks | |---|---| | mechanism | is there a route by which this could happen at all? | | magnitude | how large is it, and in whom? | | boundary | where does it stop holding — dose, age, sex, comorbidity, duration? | | alternative | what else would produce the same observation? | | harm | what does it cost, and to whom? | Three siblings from one angle are three rewordings. The angles are counted and shown. Each child records `link: decomposes` and the parent waits for them all, then rolls up into an answer. A child of a parent that is already answered records `link: continues`. **Write `link` at creation**: it cannot be worked out later. **Sharpen** instead when the direction is already narrow: one question, with the search terms you will use. A question that needs a fourth level of depth was not narrowed properly at the first. `next` gives the depth and the advisory limits (three deep, three to five wide). Go past them only for a candidate at the depth limit whose settling question cannot be written anywhere else, and say why in the body. Otherwise, at the limit, answer something instead of splitting again. ## 3. The gap sets the shape of the question - **unsearched**: nobody has looked yet. The cheapest search. - **unstudied**: we looked and found nothing. *Does it exist under any other name?* - **unexplained**: evidence is in and nothing accounts for it. *What would account for this?* (often a sign to let Explain run instead) - **unreplicated**: one independent line. *Does it replicate in a different design or population?* - **surrogate-only**: known only through a marker. *Does the marker track the outcome that matters?* - **disputed**: sources disagree and the context question did not dissolve it. *What differs between the two settings?* ## 4. Save `nodes` of `kind: question`, each with `stakes`, `angle`, the parent and `link` where there is one, and a body of one or two plain lines saying why this question. Claiming nothing has been studied needs `last_search` with two rounds in different vocabularies (refusal 3). In `expanded`, what you split and along which angles; in `open`, which children you did **not** create and why. The frontier ranks directions by moves to a verdict, so a new question starts at the bottom. Write one when the research needs it, not because the one in front of you is hard. ## Rules - **Stakes is ordinal**: `high`, `some` or `low` - how much resolving it would change the answer in chief. Never a number. - A question that cannot come out both ways is dropped, not ranked. - Rewording an existing question is a **new** question; close the old one `superseded` and cite it. Titles never change: links are made of them. - Keep at least one open direction that is not a continuation of the current best.
© Insilico-org · MIT · Insilico-org/plsi
Similar skills
All skills for Science and research →ask by other authors comes first.