idea-discovery

name	idea-discovery
description	Stage 1 of the research workflow: turn a broad direction into a validated research idea with a clear problem anchor, novelty assessment, and method thesis. Use when the user wants the idea stage only, such as brainstorming and validating research ideas.

Idea Discovery

Run the idea stage for: $ARGUMENTS

Purpose

This skill handles only the idea stage. It should end with a validated research direction, not with code, full experiments, or paper drafting.

Use this skill to go from:

broad topic -> literature map -> candidate ideas -> novelty check -> critical review -> refined top idea

Do not use this skill to:

• write experiment code
• launch training or simulation
• generate paper figures
• draft the paper

Those belong to later stages: /experiment-pipeline and /paper-writing.

Design Principles

• Kill weak ideas early. Cheap rejection is better than expensive implementation.
• Anchor the problem before polishing the method. Avoid drifting into vague "interesting" directions.
• Novelty is necessary but insufficient. The output must also be testable and worth implementing.
• Keep stage boundaries hard. End with a research-ready idea package, not an experiment plan or manuscript.

Domain Persona

Default to this lens when the project is in multi-agent formation control, cooperative guidance, distributed control, or swarm robotics:

• Think like a senior professor in multi-agent systems and formation control.
• Judge novelty as an IEEE journal reviewer, not as a workshop brainstorm partner.
• Prefer clean control-theoretic contributions, sharp assumptions, and validation that would survive IEEE reviewer scrutiny.
• Pay special attention to stability claims, communication assumptions, collision avoidance, scalability with agent count, and robustness to disturbances, delays, or graph changes.

Constants

• REVIEWER_MODEL = gpt-5.4 - Used for external critical review
• STRATEGY_MODEL = opus - Use for final selection, scope reduction, and go/no-go judgments
• EXECUTION_MODEL = sonnet - Use for literature synthesis, candidate generation, and report drafting
• AUTO_PROCEED = true - Continue across internal phases unless the user interrupts
• MAX_IDEAS = 8 - Generate a manageable candidate set, then filter aggressively
• TOP_IDEAS_FOR_DEEP_CHECK = 3 - Run full novelty and review on at most 3 ideas
• TARGET_OUTPUT = IDEA_REPORT.md - Main stage artifact

If opus or sonnet are not available in the host, keep the same split with the strongest reasoning model for strategy work and the default execution model for drafting and exploration.

Required Inputs

Accept any of the following:

1. A broad direction, problem area, or benchmark family
2. Existing notes such as IDEA_REPORT.md, STORY.md, NOTES.md
3. Optional constraints: compute budget, target venue, preferred datasets, domain focus

If the user provides nothing but a vague topic, that is sufficient to start.

Outputs

This skill should produce:

• IDEA_REPORT.md
• FINAL_PROPOSAL.md
• optional NOVELTY_CHECK.md

The outputs must be enough for /experiment-pipeline to start without redoing idea-stage work.

Workflow

Phase 1: Map the Landscape

Invoke:

/research-lit "$ARGUMENTS"

Goal:

• identify 2-4 meaningful sub-directions
• extract recurring limitations, missing evaluations, and open tensions
• separate crowded lines from under-explored ones

Checkpoint summary:

• area summary
• strongest current baselines
• likely research gaps
• recommendations on where to search for ideas

Phase 2: Generate Candidate Ideas

Invoke:

/idea-creator "$ARGUMENTS"

Generate a candidate set, then filter by:

• originality
• feasibility
• expected signal under realistic compute
• alignment with user constraints

Each surviving idea should have:

• one-sentence thesis
• target problem
• key mechanism
• expected advantage
• obvious failure modes

Phase 3: Deep Novelty Check

For the top ideas, invoke:

/novelty-check "[idea description]"

Goal:

• identify the closest prior work
• determine whether the idea is truly new, only a recombination, or already done
• record exact differentiation points

Kill any idea that is too close to existing work unless there is a clear and defensible delta.

Phase 4: External Critical Review

Invoke:

/research-review "[top idea + novelty findings]"

Goal:

• pressure-test claims
• identify the minimum convincing validation path
• surface likely reviewer objections
• reduce scope if the idea is too broad or under-specified

This stage should make the idea sharper, not bigger.

Phase 5: Refine the Chosen Idea

Invoke:

/research-refine "[top idea + novelty findings + reviewer feedback]"

Goal:

• freeze the problem anchor
• write a clear method thesis
• define the dominant contribution
• list the minimum evidence needed later in experiments

Do not produce a full experiment matrix here. That belongs to /experiment-pipeline.

The Strategy Model owns the final choice of which idea advances. The Execution Model writes the artifacts.

Output Format

Write IDEA_REPORT.md in this structure:

# Idea Report

## Executive Summary
- Recommended idea
- Why it is worth pursuing
- Main risk

## Landscape Summary
- Key sub-directions
- Gaps
- Closest baselines

## Ranked Ideas
### Idea 1 - Recommended
- Problem anchor:
- Method thesis:
- Why now:
- Novelty status:
- Main risks:
- Minimum evidence needed later:

### Idea 2 - Backup
...

## Rejected Ideas
- Idea:
- Reason rejected:

## Next Stage Handoff
- Readiness for experiment stage
- Required assets or missing prerequisites

Write FINAL_PROPOSAL.md in this structure:

# Final Proposal

## Problem Anchor
## Core Hypothesis
## Method Thesis
## Dominant Contribution
## Closest Prior Work and Delta
## Required Evidence Later
## Open Risks

Handoff Rules

A successful run must hand off to /experiment-pipeline with:

• one recommended idea
• one backup idea if possible
• explicit risks
• explicit novelty statement
• explicit minimum evidence requirements

If these are missing, the stage is incomplete.

Stop Conditions

Stop and report clearly if:

• no idea survives novelty check
• all surviving ideas are too expensive or too vague
• the problem anchor is still unstable after review

In that case, return the best salvage path:

• narrow the scope
• change benchmark family
• change contribution type
• rerun idea generation with stricter constraints

Composition

Typical next command:

/experiment-pipeline "FINAL_PROPOSAL.md"