sanad-agent: Skill for Claude Code

Jev Dual-Brain Automation

Use Jev as a fast judgment primitive inside a workflow owned by software. Jev is not an agent: it does not own phases, retries, irreversible actions, arithmetic, dates, safety policy, or completion.

Ownership model

Do not ask Jev to “complete the task,” infer a multi-step plan, calculate totals, compare exact dates, or decide whether a risky action is allowed. Ask one snap judgment per question and compose the answers in code.

Required workflow

1. Define phases before acting. Mark submit, send, purchase, delete, permission, and similar effects as irreversible.
2. Observe structured state. Include the goal, current phase, screen/page mode, relevant candidates, completed phases, and bounded action history as named JSON fields.
3. Prune candidates deterministically. Send stable IDs and concise descriptions. Never let Jev invent a selector.
4. Classify hierarchically when decisions depend on each other. First classify page/screen mode. In a later request, ask for the action among candidates valid for that mode. Questions in one request are independent; one answer cannot condition another answer in the same batch.
5. Inspect the complete answer. Choice and Score return probabilities plus confidence. Route low confidence or an invalid target to System 2 before acting. Noul is itself a yes probability and has no separate confidence field.
6. Reserve the phase before an irreversible action. A completed phase cannot re-enter. An attempted irreversible phase cannot be retried automatically after an ambiguous result.
7. Snapshot immediately before acting. Execute one exact adapter action, once.
8. Wait for events, not repeated model calls. Subscribe to navigation, lifecycle, response, or UI-transition events and block until a relevant event or timeout. Do not poll Jev while waiting.
9. Verify after the event. Use deterministic assertions for exact strings, IDs, dates, counts, and numeric values. Use a separate narrow Jev judgment only for semantic evidence.
10. Escalate fail-closed. Stop on low confidence, blockers, invalid candidates, unchanged UI, repeated targets, oscillation, timeout, or unknown state. Give System 2 bounded state, candidates, phase ledger, and recent actions.