| **English** | [简体中文](SOP.zh-CN.md)
|
|
|
| # token-saver · Manual SOP (runnable by a human, no AI required)
|
|
|
| > This is the skill's **manual operation layer**: `SKILL.md` compressed into a card a human can follow by hand — when you (a human) dispatch AI subagents manually, or want this discipline in a team runbook. The SKILL remains the rule source; this card compresses, it does not add rules.
|
|
|
| 1. **Sensitivity first**: classify the data (PUBLIC / INTERNAL / CONFIDENTIAL / RESTRICTED), filter to models allowed to touch it, and only then discuss capability and cost (details: [`references/sensitivity.md`](references/sensitivity.md)).
|
| 2. **Judge task intensity** (light execution / routine work / deep judgment / critical review) — don't go by the task's name; check five dimensions: uncertainty / cost of error / verifiability / context complexity / side effects. Any one clearly elevated → bump the tier up.
|
| 3. **Inventory on the spot**: only consult the models / reasoning options the current platform actually exposes or permits; keep no long-lived model roster.
|
| 4. **Route by interface capability**: use `EXPLICIT` when the interface lets you pin a model; `PRESET` when you can only pick predefined agents; `INHERIT_ONLY` when nothing is settable (label it honestly — never pretend a model was chosen). Log one ROUTE_DECISION record per dispatch (format: [`references/route-decision.md`](references/route-decision.md)).
|
| 5. **Cascade has preconditions**: only when output is machine-checkable and isolation / idempotency / rollback preconditions hold — run one tier lower first (never below light execution), on failed verification retry once at a higher tier, then stop and return failure evidence. Irreversible side-effect tasks (sending messages / touching databases / deploys / bulk deletes) and the critical-review tier **must not cascade**; under `INHERIT_ONLY` a rerun is a same-tier retry, never an "upgrade" (full terms: [`references/cascade.md`](references/cascade.md)).
|
| 6. **Verify after dispatch**: confirm the actual model from host run records; if the host exposes usage data, include actual spend in the flow-back (session-local reference only); mark "unverified" when you can't confirm.
|
| 7. **Models bind to task intensity, never permanently to an agent's identity.**
|
|
|
| ---
|
|
|
| > Maintained in sync with the SKILL; recheck this card whenever the SKILL version bumps. If they diverge, `SKILL.md` wins.
|
|
|
