| name | research-briefing |
| description | Reconstruct and explain a consequential research update when the user asks for a recap, has lost the thread, or needs a long delegated result, conflicting evidence, or stage transition turned into a clear decision narrative. Do not use for routine progress notes or narrow follow-ups where shared context is already established. |
Research Briefing
Produce the shortest explanation that restores the shared research state needed for the user's next judgment.
Before writing, recover the relevant ProjectView live delta, experiment record, delegated result, run identity, or artifact. Treat compacted state and assistant prose as navigation; verify a consequential factual claim against the strongest available record.
Select, rather than mechanically enumerate, the parts the user currently needs:
- orientation: where this result sits in the current stage and what question was open;
- work: what the Leader, Codex, or another actor actually changed, compared, or ran;
- evidence: the decision-carrying observation, comparison, and minimum validity check;
- interpretation: which hypothesis or route is supported, weakened, untested, or still ambiguous, including the main boundary on that claim;
- consequence: what changed in the research plan and what next question or decision now has the highest value.
Keep work, observation, validity, interpretation, and plan logically distinct even when presenting them as natural prose rather than headings. A completed task or commit is not itself evidence for a research claim.
Write for a technically capable collaborator who may not have read the intermediate transcript:
- lead with the result and why it matters;
- use explicit actors, actions, comparisons, and causal connectors;
- explain a local term or acronym once when it becomes necessary;
- attach important numbers to their metric, baseline, threshold, or uncertainty;
- use one concrete project example or compact analogy when it reduces conceptual load, then state the technical meaning;
- translate internal JSON, ledger terms, and subagent shorthand instead of forwarding them;
- omit recaps the user has already demonstrated they understand.
If ProjectView is stale, describe its structured state as the pre-result baseline and integrate the pending evidence delta before stating the current position. Do not silently promote a Codex recommendation or an unvalidated run into the project's conclusion.
