Run an Expedition

In short

You want to discover something — map an unfamiliar data set, survey a fast-moving field, probe whether an idea is feasible — and you don’t know the deliverable up front. The other guides converge on a known end; this one doesn’t. This is Michi’s learning-mode lane, the expedition, carried by /michi-expedition.

The deliverable is replaced by a portrait — an accreting picture that sharpens across many runs. The iteration cycle still runs, but within each dated run, and “done” becomes ongoing value plus a periodic readout.

A few rules for this kind of work:

• Bound the run from outside. An open mandate (“explore until exhausted”) is undecidable from the inside — set a budget and a when-to-stop before the run starts.
• The run is Entrusted; the review is Paired. The agent spirals on its own within the mandate, then surfaces a readout for you to steer.
• Surface the halt decision — don’t park silently at every turn, and don’t run on past the mandate. Bring back what’s pending, what would help, and a recommendation.

Use this pattern when the work is to discover, not to build something whose shape you already know.

A sample session

An abridged expedition session — adding a new research area to a standing “AI tooling” charter, then running its first campaign. It shows the shape: orient from disk, recognize a charter revision and settle the Paired forks before touching data, run the Entrusted pass within a mandate, then surface the halt decision with a recommendation.

You: /michi-expedition charter — I want to add a new area to the AI-tooling research: voice-driven assistants.

[The agent runs the orientation ritual — reads the structure reference, STATUS, the charter, and its portrait — before proposing anything.]

Agent: Oriented. Adding this is a charter revision — a new portrait dimension — then a campaign against it. Charters are Paired: it’s your mission, and three decisions shape everything downstream. One word each settles them:

1. Depth — a tracked special-focus theme, or just awareness-level? (Recommend: special-focus.)
2. Scope — consumer products, the underlying voice tech, or both? (Recommend: both.)
3. Campaign shape — a dedicated orient pass now, or fold it into the next weekly run? (Recommend: a dedicated pass — the portrait’s empty; it needs first blobs.)

You: Special focus. And I mean customizable, voice-driven agents I can extend with my own tools — ask questions, start research, capture notes, schedule tasks. A first-pass survey, then weekly updates.

Agent: Sharper than “voice assistants” generically — and it’s an options survey (off-the-shelf → toolkits → build-your-own), not a news tracker. A tight batch before I revise the charter and run the bootstrap:

1. Name it “Voice-Driven Agents,” measured against your use cases and your bring-your-own-tools bar. OK?
2. Scope = active web research, not just your existing feeds — they barely cover this. Good?
3. When to stop: enumerate the landscape, deep-read the ~8–12 most relevant, seed the portrait, then surface. Reasonable cap?
4. Cadence: weekly, notify-on-complete — which day?

You: Yes to all. Independent research anchored on the goals; widen it if it’s richer; weekly on Saturday.

Agent: Locked — recording the decisions. Running the bootstrap within that mandate now.

[The agent revises the charter, scaffolds the campaign and a stub portrait dimension, and runs the research as parallel passes — one per option bucket — verifying each source.]

Agent: Done — new dimension added, bootstrap run, weekly update scheduled.

tl;dr: the landscape splits cleanly into off-the-shelf, toolkit, and build-your-own.

The finding worth your attention: the closest match to your ideal may need no custom build — two off-the-shelf apps now let you attach your own tools. The one open question that decides everything: their docs don’t confirm a custom tool fires inside the voice session on a phone — a ten-minute hands-on test only you can run. It decides whether you’re done or drop to a toolkit. My recommendation: run that test before anything else. Left for you: the test, and ratifying the voice sources I proposed (I didn’t auto-add them).

One judgment call I’ll surface: I dated this as its own campaign rather than folding it into the weekly cadence, because the portrait was empty and needed first blobs. Easy to realign.

You: Good. Log the hands-on test as a priority backlog item and let the weekly run track it.

Agent: Done — logged as a priority item, linked from the portrait’s open question, and the weekly run now carries it as a standing thread. It’ll flag you only when it has something that changes the recommendation.

You: /exit

The full guide

An expedition is a sibling to the production lane, not a replacement. Where planning → session → debrief march toward a known deliverable, an expedition spirals toward clarity. The concepts — the painting metaphor, the units, the two kinds of “done” — live in Expeditions (Learning Mode); this guide is the how.

The loop

The /michi-expedition skill has five modes:

• charter — open or revise a durable mission. Charters are Paired; you can’t author someone’s mission for them.
• campaign — set the mandate for one dated run: scope, budget, when to stop, and what a finding must pass before you trust it.
• run — the Entrusted spiral: pick a thread, investigate, write a report, verify the finding, promote it.
• review — the Paired close: render a readout, fold verified findings into the portrait, decide the next move.
• exhibit — make a finding digestible when it earns a visual.

Right-size the campaign

Don’t try to boil the charter in one run. A first campaign is usually a characterize / orient pass — map the substrate before interpreting it — carved by one lens, one dataset, or one question, with the budget bounding it. The spiral widens it across later campaigns. (A design-options expedition opens differently — a goals-first ideation pass before any solutioning.)

The mandate is the leash

The mandate is what keeps the agent from either running forever or checking in every few minutes. It names the budget (passes / time / breadth), the when-to-stop (a question answered, a thread exhausted by Rule-of-3 dry passes, budget reached, or a finding that redirects the effort), and the finding-verification bar (a rung label on every claim; a null model or corroboration before a claim promotes). Confirm it before the run — that’s the Paired gate that makes Entrusted running safe.

Care calibration, not a stage gate

Crawl-walk-run asks how much care does this thread’s next step deserve? — it varies by thread. Early on a thread, lean toward “here’s what I see” (facts, coverage, gaps) and pose questions; be slow to conclude. Heavier moves — a strong claim, a null test, a materialized producer — deserve more deliberation and surfacing, because care scales with consequence.

Surface, don’t park

The failure mode is either parking silently at every boundary or running past the mandate. At a real boundary, bring it back with the decision, not just a stop: here’s what’s pending, here’s what more would help, my recommendation — continue / redirect / stop? The readout is the throttle that lets Paired review keep pace with Entrusted generation.

Paired and Entrusted

An expedition splits the modes cleanly along its two phases. The run is Entrusted — the agent spirals on its own within the mandate’s budget; that’s the point of bounding it from outside. The review is Paired, and synchronous — when you say “let’s review,” the agent takes one step and yields so you can steer, rather than pre-baking a verdict. The riskier or more novel the territory, the tighter the mandate and the more frequent the surfacing.

Pitfalls

The deliverable reflex. Building a polished producer or drawing a strong conclusion and then parking. Heavier moves deserve more deliberation and surfacing — don’t sprint to a flourish because it feels like finishing.

Running a Paired review as a solo run. “Let’s review” is synchronous: one step, then yield. Batching a long chain and delivering a finished verdict is the exact failure — it leaves you no gap to steer.

The flat pile. Reports dumped without a campaign/charter home is the anti-pattern the structure replaces. Every output has a named home — route through it before writing.

Treating the manual as read-once. Re-ground in STATUS + the charter + the mandate at the start of every run. Drift comes from skipping the orientation ritual.

---

For the lane’s concepts, see Expeditions (Learning Mode). For the skill, see /michi-expedition. Running on a non-Claude-Code harness (an isolated agent like Hermes)? See Working with Hermes.