family.skill

Got a question? Get the family first.

Multiple perspectives to help you brainstorm and research. Based on Stanford's STORM.

https://github.com/user-attachments/assets/2ae67414-e157-4779-aa61-f5096d8a164b

Install · Usage · How it works

"I don't have friends. I got family."Dominic Toretto

When to use it

• You're new to a field and want the full picture fast.
• You want to know whether a claim, or a wave of hype, holds up.
• You're about to make a call and want to know what you've missed.

For single facts, long-form writing, or picking between A and B for you, other tools fit better.

Install

Tell your agent:

Install this skill: https://github.com/xhyumiracle/family-skill

Or use the Agent Skills installer:

npx skills add xhyumiracle/family-skill -g

Or, on Claude Code, the full install:

git clone https://github.com/xhyumiracle/family-skill && ./family-skill/install.sh

The full install adds two agent definitions, member and fact-checker, with subagent spawning removed, so the headcount stays fixed. Start a new session afterwards.

Needs an agent with subagents and web search. Tested on Claude Code; Codex, Cursor, Gemini CLI and OpenCode support subagents too, untested so far.

Usage

/family <question>          standard: 5 members, one extra round on disagreements when needed
/family mini <question>     3 members, otherwise the same as standard
/family deep <question>     5 members, rounds on disagreements until the evidence stops changing, at most 3

Cost depends on the question: standard measured 480K to 900K tokens over four runs; mini and deep are not measured yet, estimated at 300K to 500K and 600K to 1.1M. Each member searches at most 15 times and stops earlier when two searches in a row find no new primary source.

Or just say it: "help me understand X from scratch", "is the hype about X real", "what am I missing about X".

You get family/<short-name>/report.md. It opens with three lines: what to do, how confident to be, and which assumption would change the answer. Below that: key findings, where the members disagree, and what nobody looked into, with sources.

How it works

flowchart LR
  Q([your question]) --> B["background brief<br/>facts and constraints only"]
  subgraph crew ["five members, each searching alone"]
    direction TB
    S1["first-hand"]
    S2["expert"]
    S3["stakes"]
    S4["dissent<br/>(the only one told your view)"]
    S5["outside view"]
  end
  B --> S1 & S2 & S3 & S4 & S5
  S1 & S2 & S3 & S4 & S5 --> V["check cited sources"]
  V --> E{"disagreement?<br/>one more round"}
  E --> R["report<br/>with sources"]

The five members split the work by where the evidence comes from, which suits questions about products, policy, history, science or personal decisions.

Perspective	Who	Basis
First-hand	the people closest to the evidence: users, participants, the team that ran the experiment; for history, the contemporary sources	argument from position to know (Walton et al., 2008)
Expert	people who can explain how and why	argument from expert opinion; the "expert" role in Critical Systems Heuristics (Ulrich, 1983)
Stakes	people who decide, pay, benefit, or bear the consequences; when a question has no stakeholders, a second research method takes this seat	the beneficiary, decision-maker and affected roles in Critical Systems Heuristics (Ulrich, 1983)
Dissent	the strongest published case against the prevailing answer in the field; works on questions with no stance	minority dissent widens a group's thinking (Nemeth, 1986); real dissent corrects better than an assigned devil's advocate (Nemeth et al., 2001); one dissenting answer reduces LLM conformity to a majority (Zhu et al., ACL 2025)
Outside view	someone from another field that has handled the same kind of problem, bringing how comparable cases usually turned out	the outside view and reference-class forecasting (Kahneman & Lovallo, 1993; Flyvbjerg, 2006); untested in AI research, so marked experimental

This is the beginning of the README. Read the rest at the original. Read the full README ↗