Trigger Word Engineering

Trigger engineering is the art of designing activation phrases that are specific enough to avoid false positives and broad enough to catch legitimate invocations. It is a tighter design space than it appears. A trigger that is too specific — "execute the CSV normalization and deduplication pipeline for Salesforce import" — will never match a natural prompt. A trigger that is too broad — "data" — will match every prompt that mentions data in any context, flooding the session with irrelevant skill loads.

The sweet spot is three to five verb-noun phrases that represent the different ways a user might express the same intent. For a CSV cleaning skill: "clean the CSV," "fix this spreadsheet," "normalize the data," "prepare for import," "deduplicate the file." These five phrases cover the linguistic variation without being so broad that they catch unrelated prompts about data analysis or data visualization.

Trigger conflicts are the most common failure mode in skill libraries with more than ten skills. Two skills both trigger on "analyze the data" — one is a statistical analysis skill and the other is a data visualization skill. Claude activates both, consuming tokens for two full skill loads when only one is relevant. The fix is intentional differentiation. Rename the triggers: "run statistical analysis" and "visualize this data." The overlap disappears.

Test triggers in a fresh session — not in an existing conversation where context might accidentally activate the skill. The testing protocol: start a new Claude Code session, type a prompt that should trigger the skill, verify it activates. Then type five prompts that should not trigger the skill and verify they do not. False negatives are annoying. False positives are destructive — they waste tokens and inject irrelevant instructions into Claude's context.