Non-Code Work
Michi isn’t only for code. The iteration cycle, assumption surfacing, and verification discipline apply whenever you’re working with an agent on sustained, multi-step work — regardless of whether the output is code or a document.
The non-code core loop
Section titled “The non-code core loop”Where code targets follow Implement → Test → Repeat, non-code targets follow:
Explore → Synthesize → Checkpoint
The agent investigates (reads files, queries APIs, analyzes data), synthesizes findings into an artifact, then checkpoints with you. Rinse and repeat until the deliverable is complete.
What’s different from code work
Section titled “What’s different from code work”Questions are sometimes output, not blockers. In code sessions, questions block progress — you need an answer to continue. In non-code sessions, discovery questions belong in the output. “How would a new user bootstrap this?” isn’t a question that blocks the agent — it’s a finding to document in the Discussion section.
Non-code agents tend to under-scope. The opposite of code scope creep. The agent produces fewer artifacts than requested — it synthesizes the obvious ones and skips the ones that require more initiative. A self-review check helps: “did I produce all the named deliverables?”
Exit criteria replace verification scenarios. Instead of Given-When-Then scenarios, define concrete exit criteria: “the evaluation document compares at least 3 options with tradeoffs for each” or “the roadmap has prioritized epics with effort estimates.”
Where Michi has been used for non-code work
Section titled “Where Michi has been used for non-code work”Knowledge work. Analyzing 1,350 Notion documents across a personal knowledge base. Four iterations of increasingly deep analysis, each correcting the previous one’s assumptions. The spiral produced a methodology: “start with the philosophy docs, not the file inventory — understanding how someone thinks should come before cataloging what they wrote.”
Infrastructure planning. Making sense of unstructured Slack guidance from a platform engineer about Kubernetes deployment. The agent decomposed five mixed directives into named items, cross-referenced each against the actual codebase state, ran OCI CLI commands to verify node image availability, and produced a two-milestone plan with categorized questions (must-ask vs. good-to-ask) for the platform engineer.
Project evaluation. Evaluating Notion database structure options for a personal assistant tool. The agent inventoried the Notion workspace via MCP tools, wrote a five-option evaluation, tracked six exit criteria explicitly, and posted to Slack when ready for human review.
Backlog building. Surveying six git repos and a Notion export to build a prioritized cross-project backlog with 30+ items, harvest candidates, and dependency maps.
Which skill to use
Section titled “Which skill to use”michi-explore is the primary non-code tool. It’s designed for investigation, research, brainstorming, and pre-planning. Flexible posture, not a checklist.
michi-session with the non-code target works for structured milestone execution when you have a plan with named deliverables and exit criteria.
michi-workshop handles quick non-code tasks with Michi discipline but minimal ceremony.
In practice, michi-explore handles most non-code work. The non-code target in michi-session is useful when the work is structured enough to have a plan doc with checkboxes.