insight-interview - SKILL.md Agent Skill

name: insight-interview description: KB-grounded Socratic interview. Searches existing notes on a topic, then runs a one-question-at-a-time dialogue to surface, sharpen, and extract your own thinking. Ends by running extract-insights on the full conversation transcript. automation: manual allowed-tools: [Bash, Read, Write, Glob, Grep, Task] user-invocable: true argument-hint: ""

Insight Interview

A Socratic dialogue grounded in your own knowledge base. It finds what you already think, then probes the edges - gaps, underdeveloped claims, contradictions, missing connections - one question at a time. Your responses feed directly into the vault at the end.

Purpose

Not to summarize a topic. Not to explain research. To find out what you actually think - and make that thinking precise enough to live in the knowledge base.

The KB is the context, not the content. Every question is anchored in a note you already wrote.

State Dependencies

Source	Location	Read	Write
Existing notes	`Brain/` via Local Brain Search	✓
Permanent notes	`Brain/02-Permanent/`	✓
Dialogue transcript	`resources/insight-interview-[slug]-[date].md`		✓
Extracted insights	`Brain/AI Extracted Notes/`		✓ (via extract-insights)

Process

Step 1: Search the KB

Run semantic search on the topic using Local Brain Search:

cd resources/local-brain-search && python search.py "[topic]" --top_k 15 --mode spreading

Also run a keyword grep across permanent notes:

grep -rl "[topic keywords]" Brain/02-Permanent/ Brain/AI\ Extracted\ Notes/ | head -20

Read the top 8-10 results. Understand:

What you already believe about this topic
Which claims are well-developed vs. sketched
Where tensions or contradictions exist between notes
What connections are asserted but not fully argued
What's notably absent (a gap where you'd expect a note)

Step 2: Map the Frontier

Before asking anything, internally map 5-7 candidate question zones:

Existing claim to probe: A strong assertion in a note that could be sharpened or challenged
Gap: An angle you'd expect him to have thought about but haven't found
Contradiction: Two notes that pull in opposite directions
Connection: A note that seems related to another topic he knows well - does he see it?
So-what: A well-documented insight with no clear practical implication
Origin: A belief stated as fact - where did it come from?
Frontier: The newest or most uncertain note - what's unresolved?

Prioritize. You'll likely cover 5-8 of these in a session.

Step 3: Open the Dialogue

Introduce with one sentence of context showing what you found in the KB, then ask the first question. Keep it grounded - reference the specific note or claim:

"You have a note [[X]] where you say '[direct quote or paraphrase]'. Given [new angle / related note / recent development], do you still hold that? Or has your view shifted?"

OR for a gap:

"You have strong thinking on [A] and [B], but I didn't find anything on [C] - which sits right between them. What's your actual take?"

One question only. Wait for the response.

Step 4: Dialogue Loop

After each response, decide:

Deepen this thread: The answer opened something - follow it (ask for evidence, for the counterargument, for the implication)
Mark and move: Useful answer, but another zone is more promising - transition to the next question
Capture and close: The response crystallized something - reflect it back precisely, confirm it's right, then move on

Question types to cycle through (from elicitation-techniques):

Type	When to use	Example
Clarification	Vague or assumed term	"What exactly do you mean by [X] here?"
Assumption excavation	Strong claim stated as fact	"What would have to be false for this not to hold?"
Counterargument	One-sided note	"What's the strongest case against this?"
Connection probe	Two related notes not linked	"Does this relate to your thinking on [Y]?"
Implication	Well-documented insight with no so-what	"If this is true, what changes?"
Origin	Belief with no cited source	"Where does this come from? Experience, reading, both?"
Precision	Fuzzy claim	"How strong is this? Always? Sometimes? Under what conditions?"

Aim for 6-10 exchanges. Stop when:

You've covered the main frontier zones
The conversation is producing diminishing returns
The user says "done", "enough", or similar

Step 5: Close

Before ending, ask one final open question:

"Is there anything on this topic that you think is important but we haven't touched?"

Give space for any loose threads.

Step 6: Save Transcript

Create a transcript file combining the full dialogue:

resources/insight-interview-[topic-slug]-[YYYY-MM-DD].md

Format:

# Insight Interview: [Topic]
**Date**: YYYY-MM-DD
**KB notes consulted**: [list of note titles]
**Questions asked**: [count]

---

**[Question 1]**
[Response]

**[Question 2]**
[Response]

...

---
*Raw transcript for extract-insights processing*

Step 7: Run Extract-Insights

Invoke the extract-insights skill on the transcript:

/extract-insights resources/insight-interview-[topic-slug]-[YYYY-MM-DD].md

This will:

Deduplicate against existing KB
Extract new insights in your voice
Create or update notes in Brain/AI Extracted Notes/
Log changes in Brain/05-Meta/Changelogs/

Report what was created or updated.

Dialogue Principles

One question at a time. Always. No lists of questions.
Show your work. Reference the specific note you're drawing from. This makes it feel like a conversation with someone who read your stuff, not a generic interview.
Don't summarize back every time. Only reflect back when something crystallized - and then be precise.
Follow energy. If a response is animated or detailed, stay in that thread.
Don't lead. The question opens space; it doesn't suggest the answer.
Tolerate "I don't know". That's a real answer. Ask what would need to be true to know.

Outputs

Dialogue transcript: resources/insight-interview-[slug]-[date].md
Extracted insight notes: Brain/AI Extracted Notes/
Changelog entry: Brain/05-Meta/Changelogs/

Usage

/insight-interview self-improving agents
/insight-interview dopamine and motivation
/insight-interview the folder paradigm