bragi-research-report

Render the Research workflow's reviewed cluster Workfiles into its final user-facing report — faithful consolidation, question-targeted structure, and concise, boundary-naming communication.

SKILL.md

GitHub

Research Report

Purpose

Render the Research workflow's reviewed cluster Workfiles into the final user-facing research report. This is consolidation and faithful communication of reviewed findings, not re-research: the report merges the reviewed streams into one answer to the research question, introduces no new claims, and is organized around the question rather than around the sources.

When to Use

  • Only when dispatched as the Research workflow's deliverable step. This skill is invoked by the dispatching orchestrator after all research streams have passed their reviews; it is not a general-purpose drafting or communication heuristic.
  • The inputs are all reviewed NN-research-cluster-<name>.md Workfiles, together with their review Workfiles.
  • The output is one concise, question-targeted report that names its own boundaries.

Workflow

  1. Read every reviewed cluster Workfile in full before writing anything — and its review, so the verdict's caveats shape what the report states. If an input's review verdict was BLOCKED — or its review is otherwise missing or incomplete — do not consolidate that material as if it were verified: either exclude it and name the exclusion in the boundaries, or flag it to the dispatching orchestrator in your reply.
  2. Identify the core answer to the research question that the reviewed findings support.
  3. Consolidate.
    • Deduplicate overlapping findings across streams.
    • Reconcile cross-stream points of contact — surface contradictions between streams, never silently resolve them.
    • Introduce no new claims: nothing enters the report that was not present in the reviewed Workfiles.
  4. Draft the report answer-first (BLUF). State the core answer within the first two sentences; organize the sections around the research question, not around the clusters or sources; make the structure scannable — headings, bullets, and tables where they aid the reader.
  5. State the four boundaries explicitly: what was covered, what was not covered, what remains uncertain, and what the user should verify externally.
  6. Calibrate length to the question, not to the input volume — a report is complete when the question is answered, not when every finding has been reproduced.
  7. Return the full report in your reply to the dispatching orchestrator, and write it to the designated Workfile path. The orchestrator relies on the content you return to carry the report forward; omitting it silently breaks the pipeline.

Quality Criteria

  • The primary answer appears within the first two sentences.
  • Every section serves the research question; no section exists to summarize a source or a cluster.
  • The structure is scannable — someone can get the gist in seconds.
  • No irrelevant information is included.
  • No reviewed finding is silently dropped — the covered / not covered / uncertain contract reflects the streams, and cross-stream contradictions are surfaced, not merged away.
  • A reviewed finding that does not serve the question is omitted as off-topic and named in the boundaries — a compact exclusion note, one line per exclusion, topic only, no elaboration; an honest exclusion, never misrepresented as uncertain or not covered; the no-silent-drop guarantee applies to findings relevant to the question.
  • No new claims appear beyond the reviewed Workfiles.
  • Length is proportional to the question's weight, not to the volume of findings.

Anti-Patterns

  • Data dump — reproducing every finding because it was reviewed, with no consolidation into an answer.
  • Buried lede — making the reader dig through background before the answer appears.
  • Source-organized report — a per-cluster or per-source digest is not a report; the report is organized around the question.
  • Re-researching — investigating the topic further instead of consolidating the reviewed findings.
  • Injecting unreviewed claims — adding conclusions, implications, or facts that were not present in the reviewed Workfiles.
  • Silently merging contradictions — resolving a disagreement between streams in the report's favor without surfacing it.
  • Boundary-free reporting — a report that does not state what was not covered and what remains uncertain overstates its own authority.
  • Padding to completeness — reproducing every reviewed finding at full length instead of answering the question.