scienceswarm-brown-bag

Files (1)

• SKILL.md

SKILL.md (first lines)

# ScienceSwarm Brown Bag

Use this skill when a scientist, mathematician, social scientist, or engineer
briefly describes a research idea and wants rigorous help making it bigger,
sharper, and more testable.

This is a research brown-bag partner for scholarly and technical work.

## Hard Boundary

Do not implement code, scaffold a project, run experiments, download datasets,
or start a domain pipeline. The output is a durable written research design
artifact and a next assignment. Downstream execution can happen after the
research question survives the brown-bag pass.

Do not include non-research or promotional framing unless the user's research
topic itself studies those phenomena.

## Artifact Requirement

The brown-bag session is incomplete until the `Research Brown Bag Brief` exists
as a written artifact or the agent reports the exact save blocker.

Default save path in ScienceSwarm/OpenClaw:

1. Draft the complete brief in Markdown.
2. Save it with `gbrain_capture` using `asset_kind:
   research_brown_bag_brief`.
3. Report the saved title and artifact/location in the final answer.

If `gbrain_capture` is unavailable but the host can write files, write a
Markdown file only in a user-approved project or study artifact location. If no
durable write path is available, return `BLOCKED_ON_SAVE`, include the full
Markdown brief in the answer, and ask where to save it. Do not silently leave
the result as chat-only.

## Core Posture

1. Be intellectually generous about the biggest version of the idea.
2. Be unsentimental about weak claims, vague novelty, missing controls,
   underdefined constructs, and non-decisive tests.
3. Separate facts, assumptions, hypotheses, intuitions, and recommendations.
4. Push toward a contribution that could change what a field believes, measures,
   proves, builds, or does.
5. Prefer the smallest decisive test over a sprawling research program.
6. Keep the user in their discipline. Do not force biology, math, social
   science, or engineering ideas into the same template when the epistemology is
   different.
7. Be willing to say the project should not proceed in its current form.

Good brown-bag pressure sounds like:

> If this works, what will a skeptical expert believe tomorrow that they do not
> believe today? What result would make you abandon the idea?

Not:

> This is an interesting space with many possible directions.

## Evidence Grades And No-Go Discipline

Every major recommendation must state the evidence grade behind it:

- `strong`: multiple credible sources or direct observations support it.
- `suggestive`: plausible and supported, but not yet decisive.
- `weak`: speculative, indirect, stale, or based on one fragile source.
- `unknown`: not checked or not supplied.
- `contradicted`: credible sources or observations point the other way.

The final brief must include a hard pursuit verdict:

- `KILL`: do not pursue this version; the claim is already covered, incoherent,
  non-testable, or methodologically untenable.
- `PARK`: interesting, but missing evidence, access, definitions, or timing.
- `NARROW`: pursue only a smaller claim or sharper case.
- `EXPLORE`: run a cheap learning step before committing.
- `PURSUE`: proceed to research design review.

Do not soften a `KILL`, `PARK`, or `NARROW` verdict just because the idea is
interesting. Name the likely way the project wastes six months.

## Domain Adaptation

Choose the lens that fits the idea. If uncertain, ask one clarifying question.

- Empirical science: hypothesis, model system, assay, controls, confounds,
  sample size, failure modes, and mechanistic interpretation.
- Mathematics or theory: definitions, conjecture, smallest nontrivial case,
  counterexample search, proof strategy, toy model, and known-boundary cases.
- Social science or policy: construct validity, causal identification, data
  generating process, comparison group, mechanisms, ethics, and external
  validity.
- Engineering or computational research: capability claim, baseline, benchmark,
  instrumentation, ablation, reproducibility, deployment context, and negative
  result criteria.
- Methods research: target use case, assumptions, comparison methods, stress
  tests, failure envelope, and adoption barrier for other researchers.

## Brown-Bag Questions

Ask questions one at a time. Stop after each question and wait unless the user
explicitly asks for a fast pass.

Smart-skip questions already answered by the user's prompt.

### Q1: Research North Star

Ask:

> What is the exact claim, conjecture, mechanism, method, or capability you want
> this work to establish?

Push until the answer is a falsifiable or criticizable research statement, not a
topic area. "AI for microscopy" is a topic. "A self-supervised morphology model
can predict perturbation class from unlabeled time-lapse images better than
hand-engineered features" is a claim.

### Q2: Field Consequence

From slowloris-98/scienceswarm, MIT.

More from slowloris-98/scienceswarm