domain-anchoring - SKILL.md Agent Skill

name: domain-anchoring description: Use when a user describes ANY research intent, goal, or topic — this is Phase 0 and must complete before any other research skill activates

Do NOT proceed to any other research skill until domain, research type, and expert persona are identified and confirmed by the user.

Domain Anchoring (Phase 0)

Overview

Bind the correct domain identity before any research work begins. Wrong expertise produces irrelevant baselines, missed evaluation criteria, and reviewer rejection.

When to Use

digraph trigger {
    "User mentions research?" [shape=diamond];
    "research-anchor.yaml exists?" [shape=diamond];
    "Confirmed by user?" [shape=diamond];
    "Invoke domain-anchoring" [shape=box];
    "Skip — already anchored" [shape=box];

    "User mentions research?" -> "research-anchor.yaml exists?" [label="yes"];
    "User mentions research?" -> "Invoke domain-anchoring" [label="no → not research"];
    "research-anchor.yaml exists?" -> "Confirmed by user?" [label="yes"];
    "research-anchor.yaml exists?" -> "Invoke domain-anchoring" [label="no"];
    "Confirmed by user?" -> "Skip — already anchored" [label="yes"];
    "Confirmed by user?" -> "Invoke domain-anchoring" [label="no"];
}

Step 1: Identify Domain and Subdomain

Extract the research domain (ML, bioinformatics, physics, chemistry, multimedia, data mining, etc.) and subdomain (e.g., few-shot learning, single-cell genomics, molecular dynamics). If ambiguous, ask the user directly.

Step 2: Classify Research Type

Determine one of four types:

Type	Focus	Typical Venue
M — Method/Model	New algorithm or architecture	ML conferences (NeurIPS, ICML, ICLR)
D — Discovery/Data	Data-driven scientific insights	Domain journals (Nature Methods, Cell)
C — Computational Tool	Pipelines, software, infrastructure	Software journals (JOSS, Bioinformatics)
H — Hybrid	New method + domain application	Cross-disciplinary venues

If type is unclear, MUST ask clarifying questions. Do not guess.

Step 3: Anchor Expert Persona

Bind the correct domain expert identity. Use this table to set reviewer_focus in the anchor file:

Domain	Reviewer Focus Areas
ML/AI	Strong baselines, ablation studies, robustness, data leakage, scaling behavior
Bioinformatics methods	Biological utility, gold-standard benchmarks, reproducibility
Biological discovery	Mechanism chains, alternative explanation exclusion, batch effects
Physics simulation	Conservation law verification, numerical stability, known-system benchmarks
Computational chemistry	Energy conservation, known-system benchmarks, force-field validation
Data mining	Overfitting checks, feature leakage, time-series future leak

Step 4: Preliminary Feasibility Check

Ask the user these questions before proceeding:

Compute resources: consumer single GPU / multi-GPU server / cluster / CPU only

Compute budget (optional): approximate GPU/CPU hours available for training and experiments (e.g., "50 GPU-hours on A100", "unlimited CPU", "8-hour daily limit"). If unknown, leave blank — it can be estimated later based on the experiment plan.

Submission deadline (optional): is there a specific conference or journal deadline? (e.g., "NeurIPS 2026, May 17"). If none, leave blank.

Data status: have data / public datasets / need to collect / unsure

Record answers in resources fields of the anchor file.

Step 5: Create research-anchor.yaml

Write docs/01_intake/research-anchor.yaml using the template at templates/research-anchor.yaml. Fill all Phase 0 fields: domain, subdomain, research_type, expert_persona, reviewer_focus, and resources. Present the completed file to the user for confirmation. Set confirmed_by_user: true only after explicit user approval.

Step 6: Transition

Once the user confirms the research-anchor.yaml, proceed to research-direction-exploration (Phase 1) in your next response — not in this one.

## ⛔ MANDATORY STOP

After presenting research-anchor.yaml for confirmation, END YOUR RESPONSE IMMEDIATELY.

Do NOT invoke research-direction-exploration or any other skill in this same response. Do NOT start literature review or any Phase 1 activity.

STOP. WAIT. The user must confirm the anchor before you do anything else.

Your final output for this phase should be the anchor summary followed by: "Please confirm the above settings. Once confirmed, I'll begin Phase 1 (Literature Review & Direction Exploration)."

Then STOP.

Anti-Pattern: Wrong Domain Expertise

Using the wrong domain lens is invisible and catastrophic. A chemistry professor doing deep learning may skip baseline comparisons; an ML researcher analyzing biological data may ignore batch effects entirely.

If domain expertise feels "close enough" — it isn't. Re-anchor.

Red Flags — STOP and Re-Anchor

Proposing experiments before anchor is confirmed
Skipping feasibility questions "to save time"
Assuming research type without user confirmation
Using generic ML reviewer lens for domain-specific work
"I'll figure out the domain as we go"
"The type is obviously M"

Common Rationalizations

Excuse	Reality
"User already told me the topic"	Topic ≠ anchored domain + type + persona. Complete the process.
"Type is obvious from context"	Obvious to you ≠ correct. Confirm with user.
"Feasibility check slows us down"	Discovering GPU constraints mid-experiment wastes weeks.
"Domain expertise doesn't matter for coding"	Wrong expertise → wrong baselines → wrong evaluation → rejection.
"I'll anchor later when it matters"	Every decision before anchoring uses the wrong lens.

Checklist

Identify domain and subdomain
Classify research type (M/D/C/H) with user confirmation
Anchor expert persona and reviewer focus
Preliminary feasibility assessment (compute, data, deadline if any)
Create research-anchor.yaml and get user confirmation
Transition to research-direction-exploration