Tom's system for making AI-built software compound: a shared team memory, a disciplined build loop, and the tools that carry work from idea to merged PR. Ten cards below teach the whole system — click a card to flip it.
Front is a question — answer it in your head before you flip. That retrieval effort is what makes it stick.
Click or press Enter to flip · printing this page lays every front and back flat.
http://localhost:7313 and name every pane out loud — PRs, needs-review, workspaces, staleness. Move on when you can say what Outpost will never do (push work at you).lore search for a topic you know well (e.g. "facilitator platform") and read three hits. Then flip cards 1, 2, and 10. Move on when you can explain where a lesson learned today lives next week.camp new warmup from the platform repo, look at what appeared in the workspace, then camp remove --force --name warmup. Move on when you can say why worktrees mean parallel batches don't collide./craft:brainstorm with a small real want. Stop when it produces a draft spec — don't build. Move on when the session has printed the exact next command to run./lore:flush, then lore sync. Flip cards 6–9. Move on when you can name what portage does that camp doesn't.camp new a workspace, start Claude inside it, and run /craft:execute. Watch the first slice land. Move on when Outpost shows your workspace and its state.Note: learning order ≠ dependency order — you meet Outpost first because it's the cheapest door in, even though the loop it observes comes first logically.
| Moment | Command | Where |
|---|---|---|
| Start a new piece of work | /craft:brainstorm | Claude session |
| Review an ADR/spec/plan | /craft:gauntlet | fresh Claude session |
| Turn a locked spec into slices | /craft:plan | fresh Claude session |
| Build it | /craft:execute | Claude session inside a camp workspace |
| Make a cross-repo workspace | camp new <slug> | terminal, from any group repo |
| Tear a workspace down | camp remove --force --name <slug> | terminal |
| Search team memory | /lore:search or lore search | Claude / terminal |
| End of every session | /lore:flush then lore sync | Claude / terminal |
| Open PRs + drive CI green | /portage:pull_request | Claude session in workspace |
| Drain small captured tasks | /ranger:refine · /ranger:execute | Claude session |
| Is the dashboard up? | trailhead outpost status / start | terminal |
The CLIs (camp, lore, portage, ranger, trailhead) work in any new VS Code integrated terminal — the shell profile loads them. The /craft:*-style commands are Claude Code plugin skills: they appear in any Claude session started after install. The one rule that changes in VS Code: a session can't leave the folder it started in, so for execute work open the camp workspace folder itself (File → Open Folder → the path camp pwd <slug> prints, under ~/.local/state/camp/levr/worktrees/) and start Claude there — or run camp new in the integrated terminal, which drops you in, and launch claude from that terminal.
The ADR — specs are point-in-time build orders; ADRs are the continuously current record of decisions. A stale spec is history, not instruction.
Small contexts stay in the smart zone: cheaper, sharper, and no bleed-over — your planning isn't polluted by brainstorm context. Each stage ends by printing the next command; you run it clean.
A camp workspace (camp new) built from the levr group — one worktree per member repo — with the Claude session started inside it.
One good shape: "It's a shared memory plus an assembly line — Claude checks what the team already knows, designs under adversarial review, builds in isolated multi-repo workspaces, and walks its own PRs to green. You spend your time on the thinking stages; the building runs itself."