Proactive workflows with clear limits
A useful automation notices a relevant change, prepares something you can inspect, and stops at its permission boundary.
The pattern below is general design guidance. Adam has verified a finite teaching pilot described in the current setup; that does not certify every example, provider, or future run.
1. Detect changes without a model
Start with ordinary code. Compare timestamps, revision IDs, or content hashes against the last successful check. Filter to approved sources, deduplicate events, and ignore unchanged items.
This detection step uses no model inference. It does not make the entire workflow free: connectors, hosting, storage, and later model work may have costs or limits.
If nothing relevant changed, stop. Avoid launching an agent just to ask whether anything happened.
2. Use optional, bounded triage
Some changes can be routed with rules alone. Where classification helps, an optional lower-cost model can sort a small, approved input into categories such as “needs a draft,” “needs human attention,” or “ignore.”
Set limits before enabling it: input size, items per run, model calls, retries, and daily usage. Treat uncertain cases as uncertain. Triage should not invent facts, assess substantive work, or decide that an obligation is complete.
Do not route private mailbox contents into a recurring cloud writing job merely because a local process is unavailable. Keep routine local-only inference local unless a separate, explicitly scoped assignment authorizes otherwise.
3. Reserve stronger models for substantive work
Use a stronger model when the task calls for reasoning, research synthesis, substantial drafting, or draft grading feedback. Pass only the necessary source material, the task's boundaries, and the required output format.
For assessment support, provide the rubric and approved evidence. The model can prepare a draft evaluation; the human retains final assessment judgment.
Require source checks. A draft should distinguish source facts, interpretations, and missing information. Include source references and retrieval dates when freshness matters. Missing access is a limitation to report, not a gap to fill with a guess.
4. Review before consequential action
Save the draft with its sources and a short verification note. A human checks accuracy, tone, permissions, and completeness before sending, publishing, changing records, or accepting an assessment.
Mark a reply obligation complete only after the authorized delivery is confirmed. A saved draft is progress, not delivery.
5. Make repeats and failures safe
Give each source change a stable identifier. Record what was processed, prevent duplicate drafts, and cap concurrent work. Use finite retries and a clear stop condition.
If access fails, usage limits are reached, or the configured model is unavailable, stop and report the dependency. Do not silently substitute a paid API, another provider, or a broader data route.
A small first experiment
Choose one non-sensitive document you control. Detect a revision with code, skip unchanged checks, and prepare a short change summary only when needed. Save it for review without sending it anywhere.
Test an unchanged document, a real revision, a duplicate event, and an access failure. Confirm that the workflow stops at its caps and leaves an inspectable result. Expand only after that bounded loop works.