moralespanitz/explore

<SUBAGENT-STOP> If you were dispatched as a subagent to execute a specific search task, skip this skill. Do the task and return structured results immediately. </SUBAGENT-STOP> <HARD-GATE> Do NOT launch any searches until the user has confirmed the topic framing. One sentence of confirmation is enough. No searches before confirmation. </HARD-GATE>

Explore Skill — Parallel + Persistent

Step 0 — Create session

Before anything else, create the session directory and open the lab notebook:

mkdir -p .research-loop/sessions/<slug>/

Where <slug> = topic lowercased, spaces to dashes, e.g. policy-compression-ai-agents.

Create .research-loop/sessions/<slug>/lab_notebook.md with:

# Lab Notebook — <topic>
Started: <date>
Status: exploring

---
## Session log

This file is the single source of truth for the entire research session. Every phase appends to it.

Step 1 — Confirm the topic

Say:

"I'll explore [topic]. Is this the right framing, or do you want to adjust it first?"

Wait for confirmation. Then append to lab_notebook.md:

## Framing
Topic: <confirmed topic>
Date: <date>
Researcher notes: <anything they said about prior knowledge>

Step 2 — Launch parallel searches

Dispatch ALL four researcher subagents simultaneously. Each task references the researcher agent definition (.claude/agents/researcher.md) and must follow its integrity commandments (no fabricated sources, URL-or-it-didn't-happen, read-before-summarize), use the numbered evidence table format (ID, Source, URL, Key claim, Type, Confidence), and apply the source quality tiers (A = peer-reviewed papers/official docs, B = reputable secondary, C = accept with caveats).

Agent 1 (researcher):
  Task: Find the top 10 most important papers on [topic].
  Use web search. For each paper return: title, year, authors, 1-sentence contribution.
  Prioritize foundational and recent (2020–2025) papers. Return a clean numbered list.
  Use the numbered evidence table format from the researcher agent definition.
  Save findings to .research-loop/sessions/<slug>/findings-papers.md.

Agent 2 (researcher):
  Task: Find active GitHub repos and benchmarks for [topic].
  Use web search. For each repo return: name, URL, stars (if available), 1-sentence description.
  Return 5–10 actively maintained repos. Use cited evidence table format.
  Save findings to .research-loop/sessions/<slug>/findings-repos.md.

Agent 3 (researcher):
  Task: Find 3 places where experts in [topic] fundamentally disagree.
  Use web search. For each debate: name it, state Side A's strongest argument, state Side B's strongest argument, explain why it matters.
  Use evidence table format with numbered sources.
  Save findings to .research-loop/sessions/<slug>/findings-debates.md.

Agent 4 (researcher):
  Task: Find explicit open problems stated in the [topic] literature.
  Use web search. Look for "future work" sections, unsolved challenges, gaps mentioned by field leaders.
  For each gap: state the problem, cite where it appears, explain why it's hard.
  Return 5–8 specific open problems. Use numbered evidence table format.
  Save findings to .research-loop/sessions/<slug>/findings-gaps.md.

Tell the user: Searching papers, repos, debates, and open problems in parallel using researcher agents...

IMPORTANT: Each dispatched subagent must follow the researcher agent's 6 integrity commandments and output format defined in .claude/agents/researcher.md. Key rules: every entry needs a URL, no fabricated sources, mark status honestly, triage before fetching full content, write findings to file progressively.

Step 3 — Save everything, show synthesis

When all 4 return, append the full results to lab_notebook.md:

## Literature (Phase 1)
<full list of papers with title, year, contribution>

## Active Repos
<list of repos>

## Field Debates
<3 debates with both sides and strongest arguments>

## Stated Open Problems
<list of gaps found in literature>

Then show the user a synthesis (not the raw dump):

"Here's the landscape in 3 sentences: [synthesis]

The 3 most interesting angles I see: A) [angle] — [why interesting] B) [angle] — [why interesting] C) [angle] — [why interesting]

Which direction do you want to go deeper on?"

Wait for response. Append their choice to lab_notebook.md:

## Researcher direction choice
Chose: <their answer>
Reasoning they gave: <anything they said>

Step 4 — Extract mental models for chosen direction

Now spawn 1 focused search:

Agent: What are the 5 core mental models every expert in [chosen direction] carries?
       Not facts — the intuitions and ways of thinking that take years to develop.

Show the models conversationally, one at a time:

"Here's the first mental model experts in this space share: [model 1 + explanation]. Does this match your intuition?"

Append all 5 to lab_notebook.md:

## Mental Models
1. [name]: [description]
2. ...

Step 5 — Generate diagnostic questions

Say:

"Let me give you 3 questions that would expose whether someone truly understands this vs. memorized it. Try answering them — every wrong answer tells us something."

Show questions one at a time. Wait for answers. For each wrong answer:

"Here's what you're missing: [explanation]"

Append to lab_notebook.md:

## Diagnostic Q&A
Q1: [question]
Researcher answer: [their answer]
Expert answer: [correct answer]
Gap identified: [what they didn't know]
...

Step 6 — Transition

Say:

"Exploration complete. Lab notebook saved to .research-loop/sessions/<slug>/lab_notebook.md

Ready to find the gaps and run the Carlini gate? → /gap or just tell me which open problem interests you most."

Subagent Reference

This skill dispatches the researcher subagent (.claude/agents/researcher.md) for parallel evidence gathering. The researcher agent definition encodes:

• 6 integrity commandments (no fabrication, URL required, read-before-summarize)
• Numbered evidence table format
• Source quality tiers (A/B/C/Reject)
• Context hygiene rules (progressively write to files, triage before fetching)
• Coverage status tracking per question

What NOT to do

• Do NOT dump all papers at once
• Do NOT skip saving to lab_notebook.md after each phase
• Do NOT advance phases without user input
• Do NOT dispatch subagents without adhering to researcher integrity commandments