Scheduled Research Refresh — Prompt Template

You are producing a research refresh for query <QUERY_ID>, topic “<ONE_SENTENCE_TOPIC>”. The idempotency key is the run timestamp; the artifact filename stem is “run-<ISO_TIMESTAMP>”. After writing the artifact, update <LATEST_SYMLINK_PATH> to point at this run. OUTPUT CONTRACT Produce a single Markdown file with exactly four sections, in this order: 1. Header line — run timestamp (ISO-8601 with offset), task name “Research refresh”, query ID, topic, prior-run timestamp (from <LATEST_SYMLINK_PATH> if it exists, else “none — this is the baseline”). 2. What changed since last run — a short, ranked list. No more than 8 sentences total across the section. Rank by relevance, not recency. If this is the baseline run (no prior), produce a one-paragraph snapshot of the topic as it stands today, 8 sentences maximum. 3. What explicitly did not change — one to three sentences naming points from the prior-run summary that are still accurate. This section prevents the “everything is new every week” bias. 4. Health footer — list the sources you searched and how many items you read per source. One line per source. If any source returned an error, say so. DIFF CONTRACT <!-- fill this in for your topic --> - Include: <SPECIFIC_INCLUDE_TYPES — e.g., “new primary-source filings, decisions, or hearings; new analyses from recognized scholars or national news outlets; official statements from named parties”> - Exclude: <SPECIFIC_EXCLUDE_TYPES — e.g., “opinion columns in regional news; reposts or syndications of items already covered last week; commentary on commentary”> - What counts as ‘new’: published or materially updated after <PRIOR_RUN_TIMESTAMP>. A piece updated since last run counts only if the update is substantive (new section, substantive correction, reversal) — not a minor edit. SOURCE CADENCE Every run, search ALL of the following. Do not shortcut by reading only the prior run’s summary. - <SOURCE_1> - <SOURCE_2> - <SOURCE_3> - <SOURCE_4> Report in section 4 how many items you read per source. ZERO-STATE BEHAVIOR If after reading all sources no item matches the Include criteria, write: “No meaningful changes since last run.” Do not invent items to fill the section. Do not elevate items that match Exclude or do not count as ‘new’ to fill space. An empty section-2 is a valid and useful output. HONESTY RAILS - Read every source every run. The presence of a prior-run summary in <LATEST_SYMLINK_PATH> does not excuse you from re-checking the sources. If you only read the prior summary and skipped the sources, the run is invalid. - Do not paraphrase the prior run in a way that makes it look like new work. New items come from this week’s reading, not from last week’s summary. - Cite each item in section 2 with source and date. No uncited claims. - If a source is unreachable, note the error in section 4 and continue with the other sources. DELIVERY Write the artifact to <ARTIFACT_PATH> (filename uses the run timestamp). Update <LATEST_SYMLINK_PATH> to point at the artifact just written. LOG Append one line to <LOG_PATH>: <ISO-8601 timestamp>, research-refresh, <query ID>, <status>, <artifact path>, $<cost>, <one-sentence message>