autopilot-charter

Autopilot Charter Skill

Inventory the reports, briefings, and reviews you produce on a rhythm, and decide — deliberately — which ones an AI should run on a schedule, which it should only draft, and which stay human.

What This Skill Produces

• A ritual inventory: every recurring artifact, its cadence, audience, and inputs
• An automate / assist / keep-manual call per ritual, with the reason
• Guardrails for each automated ritual (review gate, failure behaviour, escalation)
• A rollout order — which ritual to automate first and why

Required Inputs

Ask for (if not already provided):

• The recurring outputs the user or team produces (weekly updates, monthly reviews, monitors, digests)
• Who consumes each one and what they do with it
• Where the inputs live (git, analytics, CRM, inbox, notes) and whether an agent can reach them
• Tolerance for error per artifact — what happens if a run is wrong or missing?

Classification Framework

Score each ritual on four questions, then classify:

| Question | Points toward automating |

|---|---|

| Inputs reachable? Can an agent read the sources without a human fetching them? | Yes |

| Structure stable? Does the output look the same every cycle? | Yes |

| Cost of a bad run? Would a wrong or stale edition mislead a decision? | Low cost |

| Delta-shaped? Is the value "what changed since last time" rather than fresh judgement? | Yes |

• Automate — all four favourable. Schedule it end-to-end; the human sees the result, not the work.
• Assist — structure is stable but judgement or unreachable inputs remain. Schedule a draft; a human finishes it.
• Keep manual — high cost of error, or the ritual's value is the human thinking (performance feedback, strategy). Do not automate; record why so nobody re-litigates it.

Guardrails (required for every "Automate")

For each automated ritual, define:

• Review gate — does an edition ship unreviewed, or land as a draft for approval? Default to draft for anything audience-facing.
• Failure behaviour — if a run fails or a source is unreachable, does it skip, retry, or alert? A silent gap is worse than an error message.
• Staleness marker — every edition states when it ran and which sources it read.
• Kill criteria — what result (two wrong editions? a complaint from the audience?) takes it off autopilot.

Output Format

Automation Charter: [Team / Person]

| Ritual | Cadence | Audience | Call | Why |

|---|---|---|---|---|

| [artifact] | [weekly/monthly] | [who] | Automate / Assist / Manual | [one line] |

Guardrails for automated rituals:

[Ritual] — Review gate: [ship / draft-for-approval]. On failure: [skip+alert / retry]. Staleness marker: [where it appears]. Kill criteria: [condition].

Rollout order: Start with [ritual] because [lowest risk / most time saved]. Then [next]. Revisit this charter after [period].

Next step per ritual: use schedule-recipe to wire each "Automate" onto a runner, and delta-briefing to make recurring briefs report only what changed.

Quality Checks

• [ ] Every ritual has an explicit call — including the ones kept manual, with the reason stated
• [ ] No ritual is marked Automate with unreachable inputs ("somehow reads the dashboard" is Assist at best)
• [ ] Every Automate has all four guardrails, including kill criteria
• [ ] The rollout starts with a low-blast-radius ritual, not the board update
• [ ] The charter names who owns each automated ritual — autopilot still has a pilot

Anti-Patterns

• [ ] Do not classify everything as Automate — a charter with no keep-manual entries wasn't a decision
• [ ] Do not automate a ritual whose consumers haven't been told it's now machine-drafted
• [ ] Do not skip failure behaviour — a monitor that silently stops running is worse than no monitor
• [ ] Do not automate judgement-bearing artifacts (performance feedback, strategy calls) no matter how reachable the inputs
• [ ] Do not set a schedule tighter than the inputs actually change — a daily brief on weekly data is noise