autopilot

Conversion OS, Autopilot

A vault you update by hand dies in two weeks. Autopilot generates a battle-spec'd autonomous agent prompt from THIS vault's reality, then schedules it. One run = ingest fresh signal → file it where the routing map says → flag what needs a human → report.

Phase 0, Silent discovery (no questions)

Read: root CLAUDE.md (mode, conventions) · Company/profile.md + stack.md (the automation wishlist!) · Company/team.md (escalation candidates) · Clients/ profiles (active slugs, cadences) · Ops/pipeline.md. Cache: firm name, mode, active client slugs, candidate connectors from stack.md, escalation candidates. Summarize in ≤6 lines; ask only for corrections.

Phase 1, Ask only the gaps (one AskUserQuestion each, skip any already known)

1. Cadence, daily (recommended start) / twice daily / hourly / custom. Note honestly: local scheduled tasks need the machine awake; cloud routines/managed agents run with the laptop closed but need vault access via sync or a remote MCP (see docs/team-and-sync-guide.md). Recommend: start local daily, graduate later.
2. Connectors, probe live per references/connectors.md (list available MCP tools/integrations; match against stack.md candidates). Confirm which to wire; for each, the one or two scoping facts the fragment needs (e.g. which Slack channels, which calendar).
3. Escalation, who gets pinged when a run finds something urgent, and how (Slack DM if wired; otherwise flagged at the top of the daily note). Define "urgent": client-risk language, missed-deadline evidence, money/scope changes.
4. Budgets, per run: standard 40 reads / 25 writes / 15 transcripts / 3 escalations · light 20/10/8/1 · heavy 80/50/30/5 · custom. Hard caps, the agent stops at the cap and says so in its report.

Phase 2, Render

1. Read references/agent-prompt.md (the template) and references/connectors.md (per-connector blocks).
2. Fill every {{PLACEHOLDER}} from Phases 0–1. Splice only enabled connector blocks; delete disabled sections entirely (no dead headers).
3. Sanity pass: zero unfilled {{ }}; every client slug real; budgets present.
4. Show a preview (cadence, connectors, escalation, budgets) → confirm → save to Ops/autopilot-prompt.md (type: prompt, tags: [autopilot, generated]).

Phase 3, Schedule

Wire the saved prompt on the chosen cadence using whatever scheduler this environment provides (scheduled tasks / cron / cloud routines) with the vault root as working directory. If no scheduler is available here, give exact manual setup steps for the user's platform instead, never pretend it's scheduled. Confirm: prompt path, cadence, next run, where reports land (daily note ## Activity + ## Autopilot section).

Maintenance

Re-run /autopilot after: new connector, team change, escalation change, or when run reports show repeated budget exhaustion (raise budgets or cadence). The prompt is regenerated, never hand-patched, edits belong in vault context, which discovery reads.