research-render

name	research-render
description	Renders claims.json into a human-readable report.md for an existing research brief. Pure transformation — no new research, no new verification. Use when a developer wants to review research without re-running the full loop, or when a brief needs to be presented as a readable document. Auto-triggers on: "/research-render", "render the brief", "show me the brief", "render the research report", "I want to read the brief on X", "make the brief readable". Do not use to modify claims, re-fetch sources, or re-verify — use /deep-research for that.
allowed-tools	Read, Write

Research Render

Transform structured claims.json into a human-readable report.md.

Inputs

• <slug> — brief slug (e.g., deep-research-best-practices-2026)
• Optional: --sections <list> to include only specific sections (findings,contradictions,sources)

Preconditions

• ${CLAUDE_PLUGIN_ROOT}/../../plugins/deep-research/state/briefs/<slug>/claims.json exists
• ${CLAUDE_PLUGIN_ROOT}/../../plugins/deep-research/state/briefs/<slug>/sources.jsonl exists

If either is missing → FAIL with: "No brief found for <slug>. Run /deep-research <topic> first."

Execution

Step 1: Load

Read claims.json and sources.jsonl. Extract:

• Metadata: topic, freshness, triangulation_score, verdict, source_count
• Claims grouped by sq (sub-question)
• Unresolved contradictions
• Coverage gaps

Step 2: Render

Produce report.md with this structure:

---
topic: <slug>
generated: <today>
freshness: <from claims.json>
triangulation_score: <0..1>
verdict: READY | PARTIAL | FAIL
source_count: <N>
---

# <Topic human-readable title>

<one-paragraph synthesized answer from the claim set>

## Key findings

### <sub-question 1 human-readable label>

- <claim> [S1, S3]
- <claim, confidence: low> [S2 — single source]

### <sub-question 2 ...>

...

## Contradictions surfaced

- <claim A> [Si] vs <claim B> [Sj]. <Status: narrowed | unresolved>. <Brief explanation.>

(Or: "None.")

## Out of scope / not found

- <coverage gap>

## Sources

- [S1] <url> — <source_type>, <date or "no date">
- ...

Rules for the render:

• Every claim has an inline cite.
• Single-source claims append (confidence: low).
• Contradictions are named, not smoothed. If a contradiction was narrowed by round-2 evidence, say so.
• Write to state/briefs/<slug>/report.md.

Step 3: Report

rendered state/briefs/<slug>/report.md (<N> claims, <M> sources, verdict: <X>)

Rules

• Do NOT re-verify. Rendering is pure transformation of claims.json.
• Do NOT add claims or cites not in claims.json.
• Do NOT modify claims.json or sources.jsonl.
• If verdict is FAIL → refuse to render. Surface: "claims.json has verdict FAIL — do not render unverified briefs. Re-run /deep-research."

Failure modes

Code	Signature	Counter
F02	Added a fact not in claims.json	Every sentence in report.md ties to a claim id
F11	Smoothed a contradiction into a single narrative	Surface both sides with cites