Research explorer

Research Explorer

Overview

Research-topic exploration SKILL. Takes a broad direction, performs multi-dimensional web research with the agent's own WebSearch / WebFetch tools, and produces three structured Markdown deliverables. Single stage, full quality from the start. No Python runtime, no LLM SDK.

When to Use

• User says "I want to research X" without a specific topic.
• User wants to know "what are the hot topics in X".
• User needs help narrowing a broad field into 5–10 candidate topics.
• User asks for "research landscape overview".

When NOT to Use

• User already has a specific research question → use literature-survey or paper-writer.
• User wants a quick fact-check → use WebSearch directly.

Workflow

Step 1 — Understand the direction

Confirm with the user:

• Direction — the broad area of interest (e.g., "federated learning", "NLP for healthcare").
• Constraints — theory vs. applied, specific methods, target venue, compute budget, time horizon.
• Language — default English in conversation; reports in English unless the user requests otherwise.

Step 2 — Set up the run directory

DIRECTION="<direction>"
SLUG=$(python3 -c "import re,hashlib,sys; t=sys.argv[1]; n=re.sub(r'[\\s_]+','-',re.sub(r'[^\\w\\s-]','',t.lower().strip())).strip('-')[:40].rstrip('-'); h=hashlib.sha1(t.encode()).hexdigest()[:8]; print(f'{n}-{h}')" "$DIRECTION")
TS=$(date +%Y-%m-%d_%H%M%S)
RUN=output/research-explorer/$SLUG/$TS

mkdir -p "$RUN"
ln -sfn "$TS" "output/research-explorer/$SLUG/latest"

In commands below $RUN = output/research-explorer/<slug>/latest.

Step 3 — Multi-dimensional exploration

Run WebSearch across the following dimensions (one query per dimension, more if returns are thin):

1. Hot topics — "<direction> 2024 2025 hot topics" / "recent advances".
2. Open problems — "<direction> open problems" / "challenges".
3. Surveys — "<direction> survey 2024" / "<direction> review".
4. Benchmarks — "<direction> benchmark" / "<direction> evaluation dataset".
5. Applications — "<direction> applications" / "<direction> industry use cases".
6. Cross-field — "<direction> + <adjacent field>" (pick 1–2 adjacent fields).
7. Recent breakthroughs — papers from the last 6–12 months at top venues.

For each kept candidate, WebFetch the abstract URL to extract canonical title / authors / year / venue. Persist intermediate notes to $RUN/search_notes.md after every dimension so the work resumes cleanly.

Step 4 — Produce the three deliverables

Write these in $RUN/:

4.1 research_exploration.md

Structured analysis containing:

• Direction recap & constraints.
• Landscape map — main subfields and the relationships between them.
• 5–10 candidate topics, each with:
  • Title (specific enough to be a paper title).
  • Motivation (why this matters now).
  • Innovation angle (what would be new).
  • Feasibility score (low / medium / high) with a brief justification (data availability, compute requirements, prior work density).
  • Risk / open question.
• Recommendation — which 1–3 the user should pursue and why.

4.2 topic_matrix.md

A hierarchical Markdown outline of the topic space:

# <Direction>
## Subfield A
### Topic A.1
### Topic A.2
## Subfield B
### Topic B.1

This file is consumable by the mindmap-render skill to produce a visual mindmap.

4.3 literature_pre_survey.md

A pre-survey table of 20–30 representative works discovered above, with columns: title, authors, year, venue, URL, one-sentence relevance note. Every entry must have a URL the agent fetched in this session.

Step 5 — Optional handoff

If the user picks a topic, suggest the next skill:

• For a paper: the paper-writer skill (using the chosen topic).
• For a survey: the literature-survey skill.
• For an experiment package: the experiment-suite skill.
• For a visual topic map: the mindmap-render skill consuming topic_matrix.md.

Cross-skill data flow (path convention)

A downstream skill can locate this exploration via the slug:

• output/research-explorer/<slug>/latest/topic_matrix.md
• output/research-explorer/<slug>/latest/literature_pre_survey.md

If the user picks one topic from the matrix, downstream skills compute their own slug from the topic (not the original direction), so the slug paths diverge from this skill onward — which is correct.

Important rules

• No LLM SDK in this skill. Just a procedure + this SKILL.md.
• Candidates are suggestions, not guaranteed novel — the user must verify originality before committing.
• Feasibility scores are heuristic — flag uncertainty explicitly when relevant.
• Every literature entry must have a URL fetched in this session; no memory-only entries.