| name | research-report |
| description | Reporter-agent operation for the Deep Researcher Reflect Evolve system: produce a one-shot, PhD-level research report from an existing research plan and global research context. Use when a /deep-research session has accumulated enough findings and the user wants the final write-up generated immediately, or when the user invokes /research-report. Args: [<topic-slug>].
|
Research Report — One-Shot Reporter Agent
This skill executes Step 7 of the Deep Researcher Reflect Evolve
architecture: a single-pass synthesis of the final research report from
the persisted plan and global research context. It does not perform
new web searches — synthesis only.
When to use
- The user wants the final write-up from an existing research session
- A
/deep-research loop terminated and the user wants to regenerate
the report (e.g. with different framing) without rerunning the loop
- The user invokes
/research-report
If the working directory has empty/sparse context.md, warn the user and
suggest running /deep-research or /research-search answer … first
before generating the report.
Argument parsing
Single optional arg: a topic slug under .research/. If omitted and
exactly one .research/* directory exists, use it. If multiple, ask
the user which.
Procedure
Resolve paths from the slug:
.research/<slug>/plan.md → <plan_path>
.research/<slug>/context.md → <context_path>
.research/<slug>/report.md → <report_path>
Spawn a fresh general-purpose subagent. Brief:
You are the Report Writer in the Deep Researcher Reflect Evolve
system. Read <plan_path> and <context_path> in full. Write the
final research report to <report_path> in a single inference pass.
Requirements:
- Cohesive PhD-level narrative — integrate findings into a unified
structure, not a list of summaries per query
- Preserve every distinct fact, number, date, and source from
<context_path>. Maximize fact density.
- Cite sources inline as
[domain.com] and include a numbered
## Sources section at the end with full URLs
- Where the context flagged contradictions (
⚠ conflict:), surface
them honestly in the narrative — do not paper over them
- Length: as long as the evidence supports — do not pad and do not
truncate. PhD-level depth is the target.
- Do NOT mention the agentic process, the plan, the candidates, or
the iteration loop. The report is about the topic, full stop.
- Do NOT add a "Limitations" section unless evidence in the context
actually identified concrete limitations of the underlying topic
Suggested structure (adapt to the topic):
- Abstract / executive summary (≤200 words)
- Introduction & motivation
- Body sections corresponding to the major investigation areas
- Synthesis / cross-cutting analysis
- Conclusion
- Sources
Return when the file is written. Do not summarize what you wrote —
the file IS the deliverable.
When the subagent returns, tell the user the report is at
<report_path> and offer a one-paragraph summary.
Notes on faithfulness to the paper
- The paper deliberately uses One-Shot Report Generation rather than
TTD-DR's iterative Report-level Denoising — the reasoning is that
the high-fidelity Global Research Context built during the sequential
loop is already strong enough that one-pass synthesis preserves
computational efficiency without quality loss. This skill matches
that choice exactly.
- The paper does NOT prescribe a fixed report structure; the suggested
structure above is a sensible default the writer can adapt.