anthropic-os

Anthropic OS

<skill_contract> One owned work system with its workflow, users, traces, permissions, metrics, constraints, and review horizon. A supervised operating-system redesign with one bounded experiment, control gates, cadence, and rollback. The selected practice has a baseline, hypothesis, owner, metric, guardrail, budget, stop rule, and review receipt. <non_goals>Extreme productivity claims, surveillance, automatic policy evolution, or cadence without supporting context and capability.</non_goals>

Redesign one work system as a supervised learning loop. Plasticity means practices may change from evidence; competition means alternatives contend; constraint means attention, time, permissions, and review bandwidth shape the design. Load references/operating-system-playbook.md for diagnostics and artifacts.

Usage Template

Provide: system boundary, owner, desired outcome, users, current workflow, local metrics, traces/data, permissions, failure history, review capacity, and horizon.

Workflow

Define one operating bottleneck and baseline. Run Four-C in order: Context (truth/history), Connections (systems/accounts), Capabilities (skills/SOPs/evals), Cadence (triggers/reviews). Do not add automation cadence until the first three can support and verify it.

<unknowns_gate>

Treat productivity multipliers, culture narratives, maturity scores, and vendor case claims as hypotheses until local evidence exists. If outcome owner, trace consent, or approval authority is absent, return NEEDS_INPUT. Do not infer the expansion of local labels such as CASH when the system has not defined them.

</unknowns_gate>

1. Audit the closed loop: trace substrate, DRI, custom taste/eval, review bandwidth, prediction, feedback latency, and quiet-success stop.
2. Choose one flywheel and its bottleneck; allocate roughly 70% to a validated big bet and 30% to business-as-usual/option preservation only when local constraints justify it.
3. Apply 3B: Bending adapts a practice to context, Breaking removes a limiting rule or metric through an approved experiment, Blending combines mechanisms across domains.
4. Generate alternatives, then select one two-week-or-shorter experiment with hypothesis, owner, cohort, metric, guardrail, budget, stop, and rollback.
5. Run dual prediction when useful: human and agent predict outcome independently; compare actual result, record prediction error, and update decision weights only after repeated calibrated evidence.
6. Add a success-disaster pre-mortem: what breaks if adoption or throughput exceeds expectations; define load, quality, permission, support, and rollback controls.
7. Escalate permission through observe -> co-drive -> scoped reversible action -> monitored routine -> audited autonomy. Keys and environment enforce boundaries.
8. Review evidence; keep, adapt, retire, or combine the practice. No automatic archival or policy installation: human approval and the promotion gate govern system changes.