CoachDiscoverMapCompareSavedAI LearningAI WeeklyAbout
← Back to library
02 · Applied

Running the citation pipeline end to end

Three skills, one tracking list, one source folder — and the brief comes back fully cited. Here's exactly what happens at each step, and how to drive the model when you want to.

Takeaway 1
The list is the workspace — every skill reads from or writes to it.
Takeaway 2
Empty citation cells are the trigger — the finder only re-runs rows that need it.
Takeaway 3
Naming the skill is optional but useful when you're driving a specific process.

Create the tracking list first

Step one is provisioning the SharePoint list the rest of the pipeline writes into. The List Creator skill spins up a list with three columns: the assertion itself, the citation, and which source file the citation came from.

The list is a contract. Once it exists, the other two skills know exactly where to put their output. Without it, the extractor would have nowhere to drop records, and the citation finder would have nothing to process.

SharePoint lists are the connective tissue between skills. Each skill agrees to read from or write to specific columns — the schema is the API.

Extract every assertion as a row

With the list ready, the Assertion Extractor skill takes the brief file as input. It walks the document, identifies each standalone factual claim, and creates one list record per assertion. The citation column is left blank on purpose — that's what the next step looks for.

Open the list in a second tab while the extractor runs and hit refresh. Records appear one by one as the model works through the document.

The extractor's job is parsing, not researching. It splits the brief into atomic, citeable claims — short enough to match against a source passage, complete enough to stand alone.
It's tempting to assume the extractor already knows the sources. It doesn't — it only reads the brief. The list at this stage has assertions and empty citation cells, nothing more.

Run the citation finder against the source folder

The third skill — Citation Finder — gets pointed at the source folder and told to process every record in the list with an empty citation. For each blank row, it searches the folder, picks the best-matching passage, and writes both the citation and the source file back into the row.

This is the only step that touches the evidence library. It's also the slowest, because the model has to actually read the source documents. When it finishes, it prints a short report summarizing what it cited and what it couldn't.

Empty cells are the work queue. Run the finder again later and it'll only touch the rows that still need a citation — re-runs are cheap and idempotent.

The output: assertion plus quoted passage

The finished list pairs each assertion with the specific passage that supports it. Not just a file name — the actual sentence or paragraph quoted out, so the user can verify the match without opening the source.

From there the table becomes raw material. Export it, paste it into a footnotes section, feed it into a new document. The structured shape — one row per claim, one citation per row — is what makes those downstream uses easy.

A citation isn't the same as a search result. The finder is expected to lift the exact supporting passage, not just say "see report X." That quoted passage is what the human verifies against.

Naming the skill explicitly

The AI in SharePoint can usually figure out which skill to use from context alone. In this workflow, the user names them anyway — "use the assertion extractor skill," "use the citation finder skill" — because they're driving a very specific process and want zero ambiguity.

Explicit invocation is also good for teaching. New teammates watching the demo learn the skill names by hearing them used. Once the skills are familiar, the verbal step can drop away and the model picks them up from context.

Be explicit when you care about the path. Let the model infer when you care about the result. Both modes work — the difference is just how much steering you want to do.
Flashcards — Applied
02 · Applied
Tracking List
tap to reveal →
SharePoint list with three columns — assertion, citation, source file — that all three skills read from or write to. Acts as the pipeline's shared workspace.
← tap to flip back
02 · Applied
Assertion Extractor
tap to reveal →
Skill that reads a document and creates one list record per factual claim, leaving the citation column blank for the next step.
← tap to flip back
02 · Applied
Citation Finder
tap to reveal →
Skill that finds backing passages in the source folder for every list row with an empty citation, then writes the passage and source file back into the row.
← tap to flip back
02 · Applied
Empty Cell Queue
tap to reveal →
Rows with no citation yet — the work queue for the Citation Finder. Re-running the finder only touches blanks, making re-runs cheap and idempotent.
← tap to flip back
02 · Applied
Explicit Invocation
tap to reveal →
Naming a skill by name in the prompt instead of relying on the AI to infer it. Useful when driving a precise multi-step process or teaching teammates the skill names.
← tap to flip back