Skip to content

Agent Fundamentals

A terminal agent, a kanban board, and plain files in the repository — the complete workflow behind a real software project, from plan to App Store.

The project carries its own operating manual in the repository: a .agent/ folder holds plan, board, skills, and tools as plain markdown and JSON. The agent reads and writes these files from the terminal; the apps render the same files for the owner. One state, no synchronization, everything reviewable in a diff.

Building block Role
Agent A terminal agent (e.g. Claude Code) implements tickets, writes tools, and commits — bound by the workflow contract .agent/AGENT.md.
Plans Intent as markdown in .agent/plans/; exactly one plan is active. The agent slices it into tickets on the board (.agent/board/: Backlog → Doing → Done, append-only history per ticket).
Skills Reusable procedures in .agent/skills/ — a trigger description plus checklist-style steps with Verify: lines.
Tools Small programs in .agent/tools/ under a JSON contract (TOOL.md): schema-typed input and output, examples as machine-checked acceptance tests.
MCPs Model Context Protocol servers declared in .mcp.json — capabilities in running processes, such as the allowlisted exec server that runs project commands.

Two applications orchestrate this state:

  • iKanban AI (macOS) — the owner’s cockpit: board, plans, skills, and toolbar actions over the exec server; question protocol and ticket history included.
  • Speccify (cross-platform) — source and lifecycle management for skills and tools: expand → execute → evaluate, provenance in expansions.yaml — plus its own plans/skills/tools management, so the workflow continues on Windows, where iKanban AI does not run.

Both operate on the same repository. Remove either — the files remain valid, the project remains workable.

The tutorial

4Notice, zero to App Store: one real app built with plans, tickets, atomic commits, skills, fastlane, and App Store Connect automation.

Speccify

Expand, execute, evaluate: from a source of skills and tools to checked implementations.

iKanban AI

The file-based kanban board: plans, tickets, history, and the question protocol between owner and agent.

Practiced, not theorized

Every example on this site comes from a real project running this workflow daily — including this site.