| name | atlas |
| description | Drives an Atlas Pilot project through its lifecycle. Use this skill whenever a prompt says "Use the atlas skill" — in planning mode (interview → PRD → epics) or epic mode (implement one epic, test-first). The skill reads and writes .atlas/pilot/ and emits ATLAS sentinel lines the Atlas desktop app watches for. |
Atlas Pilot
This skill is invoked by the Atlas desktop app to drive a software project's
lifecycle. Atlas wraps this Claude Code session and observes it only by reading
the session transcript — it cannot see your screen or your tool calls' side
effects directly. Anything Atlas needs to know, you must put into the transcript
using the conventions below. Follow them exactly; the automation depends on it.
Modes
The first prompt of the session names a mode. Pick one and follow its file:
- "planning mode" → read
plan.md (in this skill directory) and follow it.
- "epic mode" → read
epic.md (in this skill directory) and follow it.
Read the matching mode file now, before doing anything else.
The sentinel protocol
A sentinel is a control line Atlas greps for in the transcript. Rules — all of
them matter:
- A sentinel MUST be the entire final line of your message. Nothing may
follow it — no punctuation, no whitespace, no closing remark.
- Emit exactly one sentinel per turn, and only when a mode file tells you to.
- Never put a sentinel inside backticks, quotes, or a code block. It must be
raw text on its own line.
- After emitting a
GATE:* or NEEDS_INPUT sentinel, stop and wait. The
next turn brings a reply — often just the word continue.
| Sentinel | Meaning |
|---|
<<ATLAS:GATE:REQS>> | Planning: requirements captured, awaiting approval |
<<ATLAS:GATE:PRD>> | Planning: PRD written, awaiting approval |
<<ATLAS:GATE:EPICS>> | Planning: epics generated, awaiting approval |
<<ATLAS:TASK_DONE>> | Epic: one task finished, ready for the next |
<<ATLAS:EPIC_DONE>> | Epic: every task finished, epic complete |
<<ATLAS:NEEDS_INPUT>> | Either mode: blocked, a human must answer |
The .atlas/pilot/ layout
All pilot state lives under .atlas/pilot/ in the repo. Ownership matters —
only write the files marked "you write":
.atlas/pilot/
project.json Atlas owns this. Do not write it.
requirements.md you write (planning, gate REQS)
prd.md you write (planning, gate PRD)
epics/NN.json you write (planning, gate EPICS; NN = 01, 02, …)
epics/NN/history.jsonl you append (epic mode)
Do not commit anything. Atlas runs git itself at epic boundaries.
Replies you will receive
continue — the gate/task was approved; proceed to the next step.
- Anything else — feedback or a modification. Apply it, then re-emit the
sentinel for the current step (planning) or fold it in and continue (epic).