Starting

Source: hooks/ways/meta/start/start.md

Frontmatter

Field	Value
description	Recognize the beginning of a working session and route to the /start skill — orient to where work was left off, run a structured interview on a greenfield repo, check the context gauge to confirm it really is session-start, then recommend planning with context already warm.
vocabulary	start begin open session kick off pick up where we left off resume orient greenfield new project fresh start what were we doing catch up state of play onboard warm context ready to work get going first thing
pattern	/start\b|let'?s (get )?start|pick(ing)? up where|where (we|i) left off|start(ing)? (the |a )?(session|work|fresh)|new (work )?session|what were we (doing|working on)
refire	0.15
macro	prepend
scope	agent
requires	['Bash(ways:*)']

The operator is opening a working session — they want to get oriented and begin, not to be handed a cold blank slate. This is what the start skill is for. Pull it up (Skill: start) rather than improvising a "what should we do?" — the skill enforces the parts that are easy to skip when you eyeball it.

start is the opening bookend to wrap. They are inverse gauge-aware siblings: both read the same instrument (ways context), but wrap confirms the session is near its end and scales the handoff down; start confirms the session is near its beginning — and if it isn't, it says so. (The symmetry is at the skill layer; start additionally carries a macro that whispers the gauge on disclosure, because it needs a proactive nudge to dissuade a mid-session start — where wrap leans on the compaction-checkpoint way firing on its own near the limit.)

Why the skill, not a freehand "where do we begin?"

Three things go wrong when the open is done by feel:

• You start in the wrong place. A repo with work in flight, a clean main, an empty directory, and no repo at all each demand a different opening move. The skill checks the actual state — git status, the TaskList and the open issues it mirrors, the CLAUDE.md where it was invoked — before proposing anything, instead of guessing.
• Greenfield gets no interview. With nothing to pick up, the temptation is to invent a plausible-looking task. The skill runs a structured interview instead — the referent of "let's start" is itself the uncertainty; name it, don't hunt for it.
• Planning starts cold. The skill gathers state first, so when it recommends planning the context is already warm — plan mode inherits an oriented situation, not a blank page. That's the whole reason the order is start-then-plan.

The gauge guard

Starting is a beginning-of-session act. If the context gauge shows the session is already half or three-quarters spent, start dissuades rather than pretends — opening fresh work deep into a used window strands it against the next compaction. The right move there is usually wrap (close and hand off) or /clear and a fresh window, not start.

The planning caveat

Like wrap and /compact, start cannot invoke plan mode for you — Claude Code forbids skills and hooks from firing / commands. So start does the orienting work and then recommends planning (or runs a lightweight planning interview inline). It prepares the ground; the operator toggles plan mode. Don't claim to have entered a mode you can't.

See Also

• start (skill) — the procedure this routes to.
• wrap(meta) — the closing bookend; same gauge, opposite pole.
• develop(meta) — once oriented, the loop start opens into.
• delivery/issues(softwaredev) — the GitHub issues the session task list mirrors (ADR-180); the continuity check leans on them.
• context-status (skill) — the gauge start reads to decide it's really session-start.