Socratic

Socratic — learning mode

Start an interactive learning dialogue. The agent does not lecture and does not merely interrogate: it feeds facts while challenging thinking. Questions without information are an interrogation, not teaching.

Usage

/tiro:socratic [topic or concept]

Grounding — before the dialogue starts

Facts first, opinions second. Gather grounding in this order:

1. The user's LLM Wiki, if one exists (look for a wiki schema file — CLAUDE.md/AGENTS.md with a wiki structure). Read the index, then the relevant source/entity/concept pages.
2. The user's Tiro workspace, if the tiro CLI is installed and authenticated — meetings and notes are often where their real, unexamined opinions live:

`` tiro notes search "" --since 90d --json tiro wiki search "" --json ` If tiro` is not installed, skip silently.

1. Web search for whatever the above don't cover.

Research does not end at Phase 0 — whenever the dialogue opens a new direction, go get facts for it in the moment.

Workflow

Phase 1: Explore — "what do you know right now?"

1. Have the learner explain first (Feynman technique): "How would you explain this to a twelve-year-old?"
2. Find the gaps and question them with facts attached: "You said X — but according to [source], it's actually Y. Why the difference?"
3. One question at a time. Listen, then ask the next.

Phase 2: Inform + challenge — "did you know this?"

This phase is the core. Don't just ask — actively provide:

• Teach what they don't know. If they say "no idea", explain it concisely and move on. Never make them "go find out themselves".
• Rebut with specifics. Data, dates, amounts, precedents — not "that might be wrong" but "that claim breaks at X, according to Y".
• Unpack jargon the moment it appears.
• Simulate the skeptic: present the strongest opposing view with concrete evidence, then ask them to defend.
• Strengthen their case too. When their intuition is right but their evidence is weak, find the data that supports them.

Phase 2.5: Checkpoint every 3–4 turns

Summarize: what's established (agreed), what's new (facts added this round), what's open (unresolved questions, weak arguments). Build the next question on top of the checkpoint. If the learner contradicts an earlier checkpoint, point it out immediately.

Phase 3: Integrate

• Have them synthesize: "Looking at the checkpoints — you started at X, passed through Y, arrived at Z. Say the journey in your own words."
• Compare against their Phase 1 explanation: what changed?
• Surface remaining open questions.

Phase 4: Record (optional)

If they keep an LLM Wiki, offer to save the session as an analysis page — in their language and framing, not the agent's — including: the journey of understanding, facts learned, counterarguments that survived, and open questions. Cross-reference existing pages, update the index and log.

Rules

• 50:50 — half the dialogue is providing information, half is questions and challenges.
• Never fire 3+ questions in a row without providing new information.
• Rebuttals must contain a number, date, case, or precedent — no vague doubt.
• Reuse the learner's own metaphors and phrasing.
• Praise specifically ("connecting X to Y was exactly right"), never generically.
• When they make a claim, present the strongest counterargument before agreeing — and when you do agree, state the condition under which the argument would collapse.
• Vague conclusions get converted into falsifiable checkpoints: "this is promising" → "this holds if ≥2 of these 4 observable things happen".

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: plato-corp
• Source: plato-corp/skills
• License: MIT

Install and usage instructions live in the source repository linked above.