Inbox processing

Summary: An automated workflow that processes incoming content — notes, captures, links, ideas — by classifying each item and routing it to the appropriate location in your knowledge system. When the classification is uncertain, it stops and asks rather than guessing. Runs on demand after a batch of captures accumulates.

What makes this automation-level: The AI handles classification and routing decisions for clear-cut items without you directing each one. It processes a batch, makes sorting decisions based on established rules, and only escalates items where the right destination is ambiguous. You review the results but don't drive each individual classification.

Map

The recurring workflow is organizing captured content. It decomposes into: scan new items in the inbox, read each item to understand what it is and where it came from, decide where it belongs (original notes, external sources, reference material, action items), move it to the destination with appropriate metadata, flag anything that doesn't fit neatly into existing categories.

The classification and routing steps were good AI candidates — they follow consistent rules and are easy to verify because you can see where everything ended up and move anything that was misclassified.

Test

Initial test was running five recent captures through a prompt with simple classification rules. The AI correctly sorted four of five — it misclassified a book highlight as original writing because the captured text had no attribution.

Adding a rule distinguishing "content you wrote" from "content you captured from elsewhere" — looking for source attribution, quote marks, or capture metadata — fixed the category confusion. The key learning: classification rules need to account for how content arrives, not just what it contains.

Integrate

Built as an on-demand workflow triggered after a batch of captures accumulates — typically at the end of a work day or during a weekly review. The AI reads each item, applies classification rules, adds appropriate metadata (date, source type, topic), and routes it to the correct folder.

The critical design decision: when confidence is low, the workflow stops and presents the ambiguous item with its best guess and reasoning, rather than silently misclassifying. This "escalate when uncertain" pattern prevents errors from accumulating unnoticed. The threshold for escalation is intentionally conservative — it's better to ask about ten items and be wrong once than to silently misfile three.

Compound

The classification rules compound through error correction. When a misclassified item is caught during review, the rule that caused the error gets refined with a more specific condition. Early versions had broad rules ("anything mentioning a book goes to sources"); over time, rules became more nuanced ("book highlights and quotes go to sources; your reflections on a book go to notes; reading lists go to reference material").

The escalation threshold also adjusts — items that were once ambiguous become clear-cut as the rules improve, so the workflow interrupts less frequently over time while maintaining accuracy. After several months of corrections, the classification runs with minimal escalations and near-perfect sorting accuracy.

Both improvements happen through the same mechanism: a misclassification or escalation is treated as a signal that the rule needs to be more specific, not as a one-off fix.

agent-use-case-evaluation — The "escalate when uncertain" pattern is a human-in-the-loop mechanism that makes automation safe for classification tasks
ai-output-verification — Classification errors are easy to catch because the user can verify results in bulk by scanning where items ended up