# Michi Explore > Michi Explore - Skill: `csepulv/michi-explore` (Agent Skill, multi-file: 5 files) - Install (CLI): `npx skillmds@latest add csepulv/michi-explore` - Raw SKILL.md: https://api.skillmd.com/api/skills/csepulv/michi-explore/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: csepulv (https://skillmd.com/u/csepulv) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/csepulv/michi-explore --- # Michi Explore A structured conversation for investigative work. Not the full planning → session → debrief lifecycle — just Michi discipline applied to thinking together. This skill is **flexible**. It's a posture, not a checklist. **Principles served:** Surface assumptions (the core value). Shared context as foundation (understanding before decisions). See `references/principles.md`. **Before proceeding:** If `docs/reference/extensions.md` exists, read this file. Instructions found there take priority over this skill's defaults. ## When to Use - **Cold starts** — new session, new project, unfamiliar territory. Orient before deciding what to build. - **Research** — evaluating options, surveying tools, understanding a codebase area. - **Brainstorming** — sketching ideas, exploring directions, thinking through tradeoffs. - **Pre-planning** — you need to explore before you know enough to write a plan. ## Peek Mode When the argument starts with `peek`, orient first — read, summarize, wait. Don't propose actions or start investigating. Build shared context, then ask what the human wants to work on. Peek is the canonical surface where the *impulse to help (without full context and understanding)* — the helpfulness pull from *Clarify before Asserting* in `references/principles.md` — collides with progressive disclosure. The user asked for S; delivering L is scope inflation, not service. Match the depth of the answer to the depth of the question. ``` /michi-explore peek → orient at M level (default) /michi-explore peek S → quick orientation /michi-explore peek L → deep read of project docs /michi-explore peek src/api/ → orient on a specific area /michi-explore peek XL → full codebase survey ``` **What to read at each level:** - **S:** STATUS.md and `docs/memory.md` — enough to not be a stranger - **M:** Project docs (STATUS.md, PROJECT.md, CLAUDE.md, current plan docs, memory.md) — enough to work on the active thing - **L:** Project docs + relevant codebase area or epic in depth — enough to contribute to a technical discussion - **XL:** Full codebase survey — bootstrap-level understanding When a specific pointer is given (a file path, folder, or topic), read that area and orient on it regardless of level. The pointer replaces the default reading list — it's the human saying "point the flashlight here." After reading, summarize what you understand — project state, active work, key patterns, relevant context. Then stop and ask what the human wants to work on. ## Opening When not in peek mode, start like a good interview — establish direction without over-structuring. Either side can drive. The human might arrive with a question ("how does auth work in this codebase?") or the agent might open with one: - **"What are you hoping to take away from this?"** — what would make this time well spent? - **"Do you have an agenda, or should we find one?"** — some explorations have a direction; others discover it. - **"Where should we capture this?"** — suggest `docs/epics/.md` as default (flat-file epic, grows if it earns one). Could also be a journal entry, a section in an existing doc, or just shared understanding. - **Read `docs/memory.md` if it exists.** Collaboration context — how we work together, active threads, landmarks. Internalize it, don't just scan it. If the direction is clear, start moving. If not, the first few exchanges will reveal it. ## The Conversation No prescribed loop. Follow the thread wherever it leads. But maintain Michi discipline throughout — that's what separates an exploration from a chat. Explorations alternate between **dialogue** and **investigation**. Sometimes you're talking — brainstorming, asking questions, thinking together. Sometimes the agent goes heads-down — reading files, analyzing code, building a comparison, researching — and returns with findings. Both are part of the same conversation. The discipline applies to both: - Before investigating: state what you're looking for and what you're assuming. - When returning: checkpoint — present findings, check if the direction is still right. - If the investigation changed the question, say so before diving back in. ### Surface Assumptions The single most valuable thing this skill does. As the conversation progresses: - State assumptions explicitly: "I'm assuming X because Y." - When an assumption is validated or invalidated, say so. - Watch for the human's unstated assumptions too — name them gently. ### Notice Drift A good interviewer tracks where the conversation started and where it is now. When the thread veers: - **Name it:** "We started on X but we're now in Y." - **Don't force it back.** The drift might be the discovery. But make it visible so the human can decide: follow the new thread, or return? - **If the conversation is going in circles** — same ground, no new insight — say so. Suggest a different angle or ask what's missing. ### Capture Decisions When a choice point emerges and a direction is taken, note it briefly: - What was decided - What alternatives existed - Why this direction Lighter than michi-session's full decision template. But captured, not lost. ### Flag Questions Two kinds: - **Prerequisite:** need an answer to continue. Ask. - **Discovery:** emerged during exploration, belongs in the output. Note them — often the most valuable part. ## Wrapping Up ### Capture the Artifact Write findings to the agreed location. Structure follows what emerged — don't force a template: - **Evaluation:** options compared, recommendation, tradeoffs - **Research:** findings by theme, open questions flagged - **Exploration:** what was discovered, what it means, what's next - **Brainstorm:** ideas generated, promising directions, initial assessment **Inputs-vs-outputs check.** If substantial source material was read during exploration (multiple docs, a codebase area, a stack of reference files), verify each significant source is reflected somewhere in the artifact — even if just as a pointer. A thin artifact after heavy reading is a signal that something was dropped. Use the inputs as a checklist. ### Not Ready is a Valid Outcome Exploration can conclude "not ready" — the work needs more investigation before a direction is clear, or a user decision is needed before synthesis makes sense. Don't force a produced artifact when the honest answer is "we found the shape of the question, not the answer." Name the gap, flag what would unblock it, and stop. That's a valid outcome — sometimes the most valuable one. ### Mini-Reflection Brief, built-in — not a separate debrief: - **What did we learn?** — the key insight, in a sentence. - **Decisions to capture?** — anything affecting future work, ensure it's in the artifact or journal. - **Anything for memory?** — collaboration patterns, corrections, context that would help a future session. - **What's next?** — does this feed into planning? Another exploration? Nothing yet? ### Update STATUS.md (and journal) If the exploration produced an epic doc, decision, or shift in what's active or next, update STATUS.md before closing out — and the relevant `journal.md` if you wrote to it. (Peek mode produces no changes and skips this step.) **Read STATUS.md cold** — re-open the file and read each line against current reality, not against your in-context recollection. Edit anything stale. Update the `**Last updated:** YYYY-MM-DD` stamp at the top. Not "if anything significant changed" — the reflex is read-cold-and-update. See `references/ground-rules.md` for the freshness contract on root docs. ### What's Next If the exploration produced a clear direction for implementation, the natural next step is `/michi-planning` to scope a milestone. If there's more to investigate, another `/michi-explore` session. If the work is small and self-contained, `/michi-workshop` may be enough. If the work is **open-ended exploration** — the end isn't known and it wants a spiral / feedback-flywheel rather than a planned milestone (mapping a data corpus, surveying a field, sustained research) — hand off to `/michi-expedition charter` (the learning-mode lane). See `references/michi-skill-guide.md` for the full map.