AI Guides › Skills & Agents
What Happens When Two Skills Disagree
By Nigel Guy · 2 min read
The first time two skills produce contradictory output for the same task,
the instinct is to treat it as a bug in whichever one you trust less, patch
the prompt, and move on. That instinct skips the actual question, which is
why the two skills thought they were doing different jobs in the first
place.
The rule: a disagreement between two skills is a specification problem,
not a quality problem — treat it as evidence that their boundaries overlap
before you treat it as evidence that one of them is wrong.
The mechanism: the Conflict Log
When two skills give you different answers to what looks like the same
request, don't silently pick the one that sounds more confident. Work
through this in order:
- Write down both outputs, verbatim, side by side. Not your summary of
them — the actual text. Differences that look like the same disagreement
from memory often turn out to be different once you can see them
together.
- Identify what each skill thinks its job is. Read the trigger
condition and the stated purpose of each, not just what it produced this
time. Most disagreements trace back to two skills quietly claiming the
same territory under different names.
- Check whether they were given the same input. A skill that expands
an abbreviation, rounds a number, or reformats a date before reasoning
over it can produce a defensible but different answer from one that
didn't — that's an input problem wearing a disagreement's clothes.
- Decide which one owns this case, in writing. Not "use your judgement
next time" — an actual rule, added to one or both skills' descriptions,
that says which of them fires when this specific situation recurs.
- Re-run the original case through the updated pair. If they still
disagree after the boundary is redrawn, the boundary wasn't the real
problem, and you're back to step one with better information.
What to skip
Skip resolving the disagreement by picking a winner once and never writing
down why — the next person (including future you) hits the same fork with
no memory of how it was settled last time. Skip merging the two skills
into one "smarter" skill as a first move; a merged skill that tries to
cover both cases is usually where the next, harder-to-spot disagreement
comes from, because now the boundary is buried inside one skill instead of
visible between two.
Guardrails
- Not every disagreement has a clean resolution. Some tasks are genuinely
ambiguous, and the honest fix is a rule that routes to a human for that
specific case, not a rule that pretends the ambiguity isn't there.
- This mechanism assumes you can actually see both skills' full output and
their descriptions. If either is a black box you can't inspect, the
actual fix is getting visibility first — a debugging habit can't work on
what it can't read.
- A boundary you draw today is a boundary you'll need to revisit as either
skill changes. Treat the Conflict Log as a living note, not a one-off
ruling.
All 751 AI guides · JulieMango plans from £17/mo