michi-expedition

The expedition skill is Michi’s learning-mode lane. Where planning → session → debrief march toward a known deliverable, an expedition spirals toward clarity — research, hypothesis, experiment, prototype, iterate, repeat. The “end” is replaced by ongoing value plus a periodic readout. Use it when the goal is to discover, not to ship something whose shape you already know. For the concepts behind it, see Expeditions (Learning Mode).

When to use

Mapping an unfamiliar data corpus — what’s in it, what’s normal, what’s anomalous
Surveying a fast-moving field and tracking how your understanding shifts over time
Probing whether an idea is feasible before committing to building it
Any open-ended work where many threads will dead-end and the deliverable isn’t known up front

If the deliverable is clear, use planning and session instead. If you’re still figuring out whether this is an expedition, start in michi-explore and hand off.

What it does

The skill has five modes, mirroring the production lifecycle but shaped for learning:

charter — open or revise a durable mission. A charter is a standing line of inquiry, revised as new data lands rather than redone. This is a Paired step — you can’t author someone’s mission for them.
campaign — set the mandate: the contract for one dated run. Scope, budget, when to stop, and what a finding must pass before it’s trusted. The mandate bounds the open-endedness from outside, which is what keeps the agent from either running forever or checking in every few minutes.
run — the Entrusted spiral: pick a thread, investigate, write a report, verify the finding against the mandate’s criteria, promote it. Every claim carries a confidence rung (read → correlate → infer → hypothesize).
review — the Paired close. Render a readout, fold verified findings into the portrait (the accreting picture), triage open threads, and decide the next move with the human. Review is synchronous — one step, then yield — not a second solo run.
exhibit — make the picture digestible when a finding earns a visual: a static chart with its underlying data, a reading guide, a tl;dr. The readout is the throttle that lets review keep pace with the spiral.

What it produces

A charter (charters/<charter>/) — the durable mission, with a target vocabulary and the gates a campaign will fix.
Campaigns (campaigns/<yyyy>/<mmm>/<date>/<charter>/) — each with a mandate, per-pass reports, follow-ups, and method learnings.
A portrait (portraits/<charter>/) — the best-current picture, accreting verified findings (including promotable negatives: what the data can’t show).
Ruminations and a backlog — the open questions (still murky) and the actionable, tracked work queue.
An iteration log — the chronological spine across all campaigns, carrying the human’s verdict per pass.

The full doc and artifact structure is defined in the expedition-structure reference.

Key things to know

The end isn’t known by design. Progress is measured by a sharpening portrait, not by a checklist burning down.
Two kinds of “done” stay distinct — surfacing (when to stop and bring it back) and finding-verification (whether a claim is trustworthy). Production fuses them; an expedition can’t.
Crawl-walk-run is care calibration, not a gate. It asks how much care a thread’s next step deserves — early on, lean toward “here’s what I see” and pose questions; be slow to conclude. Heavier moves deserve more deliberation and surfacing.
Project specifics live in extensions.md — a corpus’s field dictionary, the domain seeds, a charter’s target vocabulary. The pattern is the toolkit’s; the substrate details are local.

For the full agent instructions, see the SKILL.md source.