How do you know if your AI is actually following your style guide — or just pattern-matching?

Run these five tests to separate genuine internalization from pattern-matching:

1. Off-genre stress test
2. Unprompted rule application
3. Self-initiated deviation flagging
4. Technical-claim challenge
5. Rough-cut grammar review A pattern-matcher mirrors surface aesthetics and breaks the moment input leaves the guide's territory; an AI that holds the grammar asks questions and applies your rules to scenes the guide never covered.

The decisive check is to request something your style guide never covers — an AI that has internalized a director's grammar asks clarifying questions and applies the guide's rules to new territory, while a pattern-matcher mirrors surface aesthetics and collapses. These tests assume your guide is loaded once as persistent context; invideo is an agentic video creation tool where the invideo agent reads a treatment document once and holds it across every shot, which is what makes the results auditable.

1. Off-genre stress test. Before generating a single frame, ask for a genre or subject your guide's director never worked in and watch what comes back. In one documented production, a creator who built a 25-page James Wan director's bible requested a courtroom thriller — something Wan never made — and the invideo agent asked clarifying questions (era, nature of the threat) before building anything, then produced stylistically coherent output. Questions before frames are the signal: an internalized guide fills gaps by asking, a pattern-matcher fills them by guessing.

2. Unprompted rule application. Next, check whether the AI applies named rules from your guide to situations the guide never specifically addressed. In the same production, the invideo agent pulled a principle from page 12 of the document ('Mood Over Narrative — the substitution rule') and applied it to a scene type the document never mentioned, and separately applied a slow-shutter motion smear effect from page 17 without being prompted. Surface mimicry reproduces examples it has seen; grammar generalizes to cases it hasn't.

3. Self-initiated deviation flagging. Then generate normally and see whether the AI cross-checks its own output against the guide without being asked. In that production, the invideo agent caught shadows leaning blue-green instead of the neutral gray the document's Stage A rule specified, flagged the deviation, and offered a warmer pass — unprompted. You can make this auditable: in a separate 70-second production, the invideo agent was instructed to output 12 parameters per shot — film reference, lens, lighting plan, color script, blocking, negative prompt — so every decision could be checked against the guide line by line.

4. Technical-claim challenge. Question the AI's cinematography claims directly — lens type, aspect ratio, lighting source — and judge the correction. When challenged on a lens attribution, the invideo agent in the James Wan production corrected itself with specifics: 'Wan shoots spherical, not anamorphic. The Conjuring: 35mm, 2.40:1 hard matte. Widescreen by extraction, not optics.' A pattern-matcher either defends the error or capitulates vaguely; a grounded system corrects with verifiable detail, like the 85:15 dark-to-light ratio that guide encoded for lighting language.

5. Rough-cut grammar review. Finally, send the assembled draft back with an open-ended 'what's working, what's not' prompt and check whether the feedback references your guide's framework rather than generic notes. In the documented case, the invideo agent caught that the entity's reveal shot was running at the wrong emotional stage register — Stage D instead of Stage C per the document's five-stage structure — a structural judgment the director had missed. Scoring a cut against your guide's named registers is something surface-level matching cannot do.

If your AI fails these tests, the usual cause is context dilution — rules losing weight as a long conversation grows — which is a setup problem to fix in how you load the guide, not a diagnostic one.

These tests are some of the ways to probe this — what convinces you will depend on how detailed your own style guide is.

Watch some of these to see what works for you:

Before generating a single frame, I stress-tested the doc. I asked for a courtroom thriller through the James Wan lens. Something he's never made. If the agent was just mirroring style superficially, it would fail here.

— invideo's creative team, documenting a James Wan-style production