philset

Go from ignorance to vague familiarity across a whole name-dropping domain — "just deep enough to hurt." The breadth-first sibling to /study: point it at a domain you don't know (often just a name), and it researches the terms worth recognizing, teaches them in batched 1-2-paragraph lessons with a quiz per item, and leaves a durable one-line-each summary you re-skim before the room. Recognition, not reconstruction. Replaces "three hours on Google." Distinct from /study (deep, one system, reconstruct-under-pressure) and /refresh (re-warm what you already learned). Usually user-invoked; agent may suggest it when "make sure I recognize these next time" intent appears.

Assess the current state of a feature, system, or area. Produces a durable assessment doc that downstream skills (/draft, /plan) can consume. Use --capacity to include napkin math (QPS, storage, bandwidth).

Create a working design document for collaborative iteration. Use when scoping a feature, system change, or workflow — the doc is the conversation artifact, not a summary.

Session startup. Walks the .meta/ tree to load user and domain context, orients to the current project, and gives a brief status summary so we can hit the ground running.

Lightweight session start — the sparse floor below /hello. Does a minimal tree walk (signpost-flag inheritance + root WORKFLOW.md) then reads the local .meta/ — but no full state read and no network. Auto-escalates to /hello for domain/neighborhood/full-state context; never runs MCP/API/calendar or hello.check. Use when you want to get moving fast without the full orientation.

Pre-merge review. Diffs the session's work against the base branch, resolves its review dimensions (configurable via signpost `review.dimensions`/`extra-dimensions`, else inferred from the medium — code default: bugs/efficiency/redundancy/architecture — with `redundancy` the recommended-but-configurable example), always runs the structural dimensions (design/track/merge reconciliation), then presents fixes for approval before committing.

Session retrospective and calibration. Use mid-session (1-2 questions) when you hit a specific friction point or made a judgment call you're unsure about. User invokes with /retro for a fuller end-of-session retrospective.

Lightweight iteration mode for grab-bag branches. Creates a track file, enforces note-before-code for non-trivial changes, and uses /defer as the scope-escalation escape hatch. Always user-invoked.

Deeply learn a system that already exists — a codebase, an architecture, a primary source, or your own recent work — by reading the actual source in stages with a write→annotate→comment→quiz→score loop. Produces a durable study doc with retention built in. Use for onboarding, interview prep, primary-source mastery, or defending your own work. Distinct from /assess (snapshot our system) and /draft (design something new). Usually user-invoked; agent may suggest it when deep-learning intent appears.

Make a scoped addition to an accepted design without superseding it. Use when an accepted design's reasoning needs updating but its decision and structure still stand. Agent-invoked when amendment intent is recognized.

Capture future work with automatic provenance and route it to the right project's roadmap or inbox. Use when scope exceeds the current session, or to build backlog from conversation. Agent-invoked when deferral intent is recognized.

Bless a design doc and begin implementation. Marks the design as accepted, archives superseded docs, updates state files, then starts building.

Process inbox todo items into the curated roadmap (or resolve them). The triage gate between staging and committed backlog. Use when todo.md has accumulated and needs sorting. Always user-invoked.

End-of-session housekeeping. Updates decisions.md and in-progress.md for the current project so the next session can pick up cleanly without --resume.