Research, plan, build, and share

This is Plan Desk’s full arc, end to end: a fuzzy idea becomes a researched plan, the plan becomes the single source of truth your agent builds from, and the live plan becomes a window your team or client can watch — and feed back into. Four moves, one connected graph, all local-first.

Each move has a hands-on guide; this page is the map that connects them.

1. Research & plan with Claude Code

Start with an idea and let Claude Code do the thinking on the page. Hand it the goal and have it research the codebase, then scaffold a plan in one call — scaffold_project_from_plan creates the project, a task per unit of work, the dependency edges between them, and a linked spec, atomically. You review the graph on the Flow canvas and steer with comments instead of rewriting the brief.

The plan is the brief: imperative task labels, concrete specs, and real dependency edges (blocks, depends_on, feeds, …) so what’s unblocked is never a guess.

→ Hands-on: From idea to development with Claude Code · concepts: Plan & execute a project

2. Track with Plan Desk

Once the plan exists, it’s the live source of truth — not a doc that rots. Tasks carry status (scope → todo → in_progress → done), specs live on the nodes, and every view (Flow, Board, Agents activity, document comments) updates over SSE with no refresh. People and agents read and write the same graph, locally, offline-capable.

Bind your repo so every session resolves the project automatically:

plandesk connect --project "Add rate limiting to our public API"

This writes the project binding plus .plandesk/skill.md — the conventions your agent follows (the build loop, dependency vocabulary, the comments workflow). See The Skill.

3. Build with live tracking

Claude Code works the plan as a loop, in dependency order: get_next_task (the next todo whose prerequisites are all done) → read the linked spec → make the smallest correct change → run tests → update_task to in_progress, then done → record_agent_progress. Because dependencies are real edges, it never starts blocked work.

You watch it happen live and steer mid-build with comments — drop a note on a spec and the agent picks it up with list_comments, adjusts, and resolve_comments it, live in your UI.

→ Hands-on: From idea to development with Claude Code

4. Share with your team

When you want a client or another team to see progress — and feed issues back — share a read-only, live projection of the plan. They open a link, join by name, and watch status move; with submit enabled they file issues into a moderated inbox. You pull those in, triage them into real tasks, and your agent builds them — status flowing back to them live. Your repo, internal docs, and edit access never leave your machine: the projection is allow-list and participants can only propose, never write.

→ Hands-on: Plan → share → build with your team · architecture: Collaboration & sync

The whole loop

Move	What happens	Driven by
Research	Idea → researched plan on the canvas	Claude Code + scaffold_project_from_plan
Track	The graph becomes the live source of truth	Plan Desk (local, SSE-live) + plandesk connect
Build	Agent builds in dependency order; you steer with comments	get_next_task loop + comments
Share	Team watches live, files issues; you triage into tasks	plandesk deploy / share create / pull + MCP triage

Same connected plan throughout — no handoffs, no status decks, no copy-paste between tools.