AI Guides › Skills & Agents

The Difference Between A Skill's Trigger And Its Actual Job

By Nigel Guy · 2 min read

When a skill misfires — runs when it shouldn't, or doesn't run when it should — the usual fix is to reword the trigger. Add a keyword, tighten a phrase, exclude a case. That fix works for exactly one misfire and quietly breaks down the moment the request is phrased even slightly differently, because the trigger was never the actual problem.

The mechanism: the Trigger/Job Split

Write these as two separate sentences, not one:

1. The trigger sentence. "This fires when the request contains X" or "when the situation looks like Y." This is pattern-matching. It can be loose, it can miss edge cases, and it's meant to be revised often as you see new phrasings come in.
2. The job sentence. "This exists to do Z, and only Z." This is the actual scope of work. It should barely ever change, because it's the reason the skill exists at all.

Once you have both written down separately, a misfire sorts cleanly into one of two buckets:

• Trigger problem: the skill fired (or didn't) on a request that, on reflection, clearly is (or isn't) an instance of the job. Fix: adjust the trigger wording. The job stays the same.
• Job problem: the skill fired correctly, but what it then did wasn't actually what was needed. Fix: the job description itself was too broad, too narrow, or wrong — and no amount of trigger tuning will help, because the trigger was never broken.

Most skill maintenance time gets spent treating job problems as trigger problems, which is why some skills accumulate an ever-longer list of keyword exceptions and never actually get more reliable.

What to skip

Skip tuning a trigger to cover a one-off phrasing you saw exactly once — that's the kind of overfitting that makes the trigger sentence unreadable within a month and doesn't generalise to the next odd phrasing anyway. Skip rewriting the job description in response to a single misfire before you've checked which bucket it's actually in; job descriptions should change on a pattern of evidence, not on one bad run.

Guardrails

• Some misfires are neither: the request was genuinely ambiguous, and no trigger wording resolves it cleanly. That's a case for a clarifying question in the skill itself, not a tighter trigger.
• A trigger you loosen to catch more real cases will also catch more false ones. There's no setting that eliminates both; you're choosing which error you'd rather have.
• This split doesn't fix a skill whose job was never well defined to begin with — if you can't write the job sentence in one clean line, that's the thing to fix first.

All 751 AI guides · JulieMango plans from £17/mo