Three example prompts trigger the three paths: column-only, view-only, and both. Each prompt walks the same branching skill — only the loaded reference file changes — and finishes with a styled list.
Meta takeaways
Signal
The prompt itself is the routing signal — keywords like "column" or "view" tip the decision.
Feedback
Progress bars in the run output reveal which branch the skill took.
Reuse
Brand and governance overlays use this exact shape — heavy file, load on demand.
The classifier reads the prompt before doing the work
When a request comes in, the parent skill looks at the wording and decides which branch to take. The video shows this clearly in the run output: the skill announces it is "targeting the column formatting data set," then loads the column patterns file, then does the formatting.
The classifier step is observable. You can tell which branch the skill picked because it logs the decision before reading the reference file.
This makes debugging fast. If you wanted view styling but got column styling, you can see the misclassification in the trace and adjust the prompt.
Prompt 1 — column path only
The first request: add color coding to my status column. The "status column" keyword signals column formatting. The classifier picks the column branch, loads the column-patterns reference, and applies a coloured status indicator to that one column.
A narrow, columnar request loads only the column reference. The view-formatting file never enters context for this turn.
This is the cheap case. The skill skips all the view-formatting schema entirely.
Prompt 2 — view path only
The second request targets a row-level treatment — the kind of thing view formatting is for. The classifier picks the view branch, loads the view-patterns reference, and applies a view-level layout. The column-patterns file stays unloaded.
View formatting isn't only for layouts like timelines and cards. It also covers row highlighting, conditional row backgrounds, and anything that styles a whole row at once.
Prompt 3 — both paths together
The third request: make my list beautiful, highlight overdue rows in red, and format the priority. "Highlight overdue rows" is view formatting (per-row). "Format the priority" is column formatting (per-cell). The classifier picks both branches and loads both reference files in the same turn.
The classifier doesn't have to pick exactly one branch. A generic "make this beautiful" request triggers a multi-branch load when the request implicitly needs both.
The author then runs a follow-up: make this list beautiful, and stick to our forest brand. This still triggers the both-paths branch — and demonstrates a second skill (brand definition) layered on top.
The "both" branch isn't slower than picking one. Once you've decided to load two files, the rest of the formatting work is the same single pass.
Why this pattern travels well
The author calls out the general principle near the end of the clip. Anywhere you have heavy knowledge that you only sometimes need — governance rules, naming conventions, brand definitions, security policies — the branching pattern fits.
A branching skill is a way to keep your AI environment lean by default and rich on demand. Heavy reference files stay one classifier away, not always-on.
The same parent skill can grow new branches as your team writes new reference files. The classifier gets a new condition; the existing branches don't change.
Flashcards — Applied
02 · Applied
Classifier step
tap to reveal →
The first thing the parent skill does — read the prompt, pick which reference file(s) apply, log the decision, then load.
← tap to flip back
02 · Applied
Sub-skill file
tap to reveal →
A reference file behind one branch — the column-patterns file or the view-patterns file. Heavy and detailed, loaded conditionally.
← tap to flip back
02 · Applied
Brand overlay
tap to reveal →
A second skill that layers brand rules (the "forest brand") on top of the formatting work, so styled lists stay on-brand.
← tap to flip back
02 · Applied
Governance pattern
tap to reveal →
The general use of branching skills to gate heavy knowledge — governance, security, naming rules — behind a classifier that loads on demand.
← tap to flip back
02 · Applied
Progress bar trace
tap to reveal →
The run output shown by SharePoint AI as the skill executes — surfaces the classification and lets you confirm which branch was taken.