Sparse priming
Turn a concept-heavy source into a compact list of meaningful cues.
Render the supplied information as a sparse priming representation: a compact list of statements and cues another language model can interpret. Keep the main concepts, relationships, distinctions, and dependencies. Use succinct assertions, associations, and existing analogies when they convey the source efficiently. Prefer explicit relationships to isolated keywords when the connection would otherwise be ambiguous. Remove repetition and rhetorical framing. Preserve proper names, important numbers, negation, uncertainty, and exceptions. Include enough context to identify what each cue refers to. Do not invent associations, facts, or metaphors to fill gaps. Do not assume a later model can reconstruct omitted details. If exact wording is required for code or a quotation, preserve it. Return only the cue list, with no explanation of the technique.
Example
Fictional retrieval-system context
Before 123 words
The Beacon document index contains manuals from the North and South warehouses, but it does not contain customer records. Each indexed passage stores its source file and page number so that a response can point back to the manual. Manuals are reindexed every night at 02:00 UTC. A manual uploaded after that run will not be searchable until the next successful run. If a run fails, the previous index stays available and an alert is sent to the operations inbox. This means availability of the old index does not establish that the latest manual is present. The team suspects large PDF files contributed to Tuesday's failure, but it has not established the cause. The description repeats that source references help readers check answers.
After 67 words
- Beacon index: North and South warehouse manuals; excludes customer records. - Each passage → source file + page number → checkable manual reference. - Reindex: nightly, 02:00 UTC. - Upload after run → searchable only after next successful run. - Failed run → previous index remains available + operations inbox alert. - Available index ≠ latest manual included. - Tuesday failure: large PDFs suspected; cause unconfirmed.
Examples illustrate the method. They do not measure model output.
Files and sources
- prompt.txt Complete prompt as plain text.
- SKILL.md The same prompt with a skill header.
- llms.txt (JSON) Prompt, metadata, example, and sources.
- LICENSE MIT license text.
Adapted from daveshap/SparsePrimingRepresentations · Sparse priming. Source notes.