Community

game-day-facilitation Agent Skill

Open SKILL.md package · productivity · Claude Code, Codex, Cursor, OpenCode

Facilitate a game day or on-call shadow with objectives, injects, abort criteria, observer roles, and a blameless debrief. Use whenever the user mentions game day, gameday, on-call shadowing, incident drill facilitation, tabletop exercise, or "run a game day" — even if they also mention chaos, failover, paging, or incident response. Do not use for a live SEV or a chaos-experiment hypothesis.

View on GitHub

Game Day Facilitation

A game day is a scheduled drill of people and process. Produce a facilitation pack: objectives, roles, inject script, abort criteria, observer sheet, and debrief. Fault hypotheses belong to chaos-experiment-design. A live SEV belongs to incident-command.

Workflow

  1. Objective. One learning goal (runbook, paging, failover comms, shadow independence). "Break prod and see" is not an objective.
  2. Scope and env. Staging, or an announced approved prod window. Named blast radius. Named abort conditions before any inject.
  3. Roles. Facilitator, inject operator, primary, shadow, observers, scribe. Dual-hat on a small team is fine; say who wears which hats.
  4. Inject script. Timed cards. Each inject has stop/rollback. No surprise extras mid-drill.
  5. Safety brief. Blameless. How to abort. Who may abort. Freeze if a real incident starts.
  6. Run. Facilitator owns the clock. Observers stay quiet on the bridge unless safety is at risk.
  7. Debrief same day. Facts, process and runbook gaps, 1–3 owned actions. No blame.

Roles

  • Facilitator — clock, inject cadence, abort, psychological safety.
  • Inject operator — executes the scripted fault; never freelances.
  • Primary — treats it as real; owns mitigate decisions.
  • Shadow — does the work or narrates; primary can take over.
  • Observers — watch comms, decisions, runbook use; no coaching unless the facilitator asks.
  • Scribe — UTC timeline of detections, decisions, injects, aborts.

Output format

# Game day: <name>
**When / env:** …
**Objective (one line):** …
**In scope / out of scope:** …
**Abort if:** …
**Who may abort:** …

## Roles
| Role | Name | Notes |

## Inject script
| T+ | Inject | Expected | Rollback | Owner |

## Observer prompts
- Comms: …
- Decisions: …
- Runbook: …
- Shadow: …

## Safety brief (read aloud)
- This is a drill. Blameless. Freeze on real SEV → `incident-command`.
- Never surprise prod. Abort is success, not failure.

## Debrief
**What we saw:** …
**Runbook / paging gaps:** …
**Actions (max 3):**
- [ ] … — Owner: … — Due: …

Rules

  1. Never surprise production. Refuse unannounced prod injects; offer a scheduled drill or a tabletop instead.
  2. Abort criteria are mandatory and specific (error budget, customer impact, real pager, facilitator call). Abort ends the drill cleanly.
  3. Psychological safety: no gotchas, no public scoring of people, no blame in notes or debrief.
  4. One inject variable at a time unless the script already says otherwise.
  5. Do not invent SLO numbers, customer counts, or tool syntax.
  6. Fault hypothesis / blast radius → chaos-experiment-design. Keep this skill on roles, injects, observers, and debrief.
  7. Live SEV → stop the drill and hand to incident-command.

Edge cases

  • Tabletop only: same pack; injects are narrative cards, not live faults. Still abort, observers, debrief.
  • On-call shadow, no inject: objective is the pager path and handoff; observer watches; still debrief.
  • Real incident during the drill: abort immediately; announce freeze; remaining work is incident-command.
  • "Kill pods in prod for fun": refuse. Staging or a scheduled, approved window with abort criteria only.
  • User wants only the chaos hypothesis: point at chaos-experiment-design; do not turn this into experiment design.