Skip to content

openJiuwen-ai/sciencediscovery / skills/report-writer

report-writer

Verified

Receives Domain Summaries (Knowledge Summary, Analysis Summary, or both) from the science-research-team orchestrator, performs cross-domain synthesis (identifying SUPPORTS/CONTRADICTS/INFORMS/INDEPENDENT connections), surfaces contradictions verbatim, and produces a formatted output per user requirements (or the default 7-section template if no format specified). NOT for performing research, conducting analysis, evaluating results, or resolving contradictions.

Report Writer Skill

Overview

This skill synthesizes Domain Summaries into a formatted output. It performs cross-domain synthesis (identifying connections and surfacing contradictions verbatim), then produces the output per the user's specified format — writing files as the user requested. If no format is specified, it defaults to the 7-section template. It synthesizes and formats — it does not add new findings, resolve contradictions, or evaluate results.

Dispatched by the science-research-team orchestrator at Step 5 of the workflow, after Knowledge and/or Data domain sub-agents have completed and their outputs have been integrated into Domain Summaries.

When to Use This Skill

Use this skill when:

  • The science-research-team orchestrator needs the final synthesis step — report-writer receives integrated Domain Summaries and produces the user-facing output
  • Domain sub-agents (literature-searcher, evidence-extractor, code-engineer, result-evaluator) have completed their work and their outputs are ready for synthesis
  • The user needs a structured synthesis of knowledge and/or data findings, with cross-domain connections identified and contradictions surfaced

Do NOT use this skill for:

  • Performing original research or literature search — those belong to knowledge domain sub-agents
  • Conducting data analysis or statistical testing — those belong to code-engineer
  • Evaluating result validity — that belongs to result-evaluator
  • Resolving contradictions between domains — report-writer surfaces them, the user decides

Input Sources

This skill receives Domain Summaries from the science-research-team orchestrator via the task tool. Input format:

### Domain Summaries
[Path to `knowledge_summary.md` if knowledge domain was activated, path to `analysis_summary.md` if data domain was activated.]

### Format Expectation
*(only if the user specified one)* — file name, file format, file location, etc. that the user requested. Omit if not specified.

### Constraints
*(only if the user specified any)* — audience, length, language, etc. that the user requested. Omit if not specified.

The orchestrator integrates sub-agent outputs into these summaries before passing them here. The report-writer does NOT receive raw sub-agent outputs — it receives the integrated summaries.

Knowledge Summary format (from orchestrator integration):

### Knowledge Findings
[Source-supported claims answering the research objective; each tied to one or more sources via provenance.]

### Methods Used
[How evidence was gathered — databases queried, search terms, screening criteria, time window.]

### Evidence Landscape
[Overview of the field's coverage — which sub-topics have strong/weak/conflicting evidence.]

### Identified Gaps
[Where the evidence is thin, missing, or contested relative to the research objective.]

Analysis Summary format (from orchestrator integration):

### Objective
[The specific analysis objective this Summary addresses.]

### Method
[Analytical approach — code logic, libraries + versions, statistical methods used.]

### Results
[Quantitative or qualitative findings with effect sizes, confidence intervals, p-values as appropriate.]

### Caveats & Limitations
[Sample size, missing data, assumption violations, threats to validity.]

### Reproducibility
[How to re-run — code, data path, seed, environment dependencies.]

Minimum input: one Domain Summary (Knowledge or Analysis). If neither is provided, cannot synthesize — request re-dispatch.

Python Package Installation

If you need to install new Python packages, install them through the Tsinghua PyPI mirror for reliability:

pip install [python package] -i https://pypi.tuna.tsinghua.edu.cn/simple

Workflow

Step 1: Read and Validate Domain Summaries

  1. Read the Domain Summary files at the paths provided in the input
  2. Determine which domains are active:
ScenarioAvailable inputsSections affected
Dual-domainKnowledge Summary + Analysis SummaryAll 7 sections populated; Sections 4–5 contain cross-domain connections and contradictions
Knowledge onlyKnowledge Summary onlySection 2 populated; Section 3 notes [DATA DOMAIN NOT ACTIVATED]; Sections 4–5 note [SINGLE DOMAIN]
Analysis onlyAnalysis Summary onlySection 3 populated; Section 2 notes [KNOWLEDGE DOMAIN NOT ACTIVATED]; Sections 4–5 note [SINGLE DOMAIN]
  1. Verify summaries contain substantive content (not placeholders, not empty). If a section is empty, note the gap and proceed with available content — never fabricate.

Step 2: Execute 4-Phase Synthesis Protocol

Phase 1 — Domain Classification and Output Format Determination

  • Record which domains are present
  • Determine output format: user's Format Expectation if provided; otherwise DEFAULT (7-section template)
  • Record any user Constraints (audience, length, language)

Phase 2 — Cross-Domain Synthesis

Extract key findings per domain — read each Domain Summary's sections:

  • Knowledge Summary: Knowledge Findings → Section 2.1; Evidence Landscape → Section 2.2; Identified Gaps → Section 2.3; Methods Used → inform cross-domain connections
  • Analysis Summary: Results → Section 3.1; Caveats & Limitations → Section 3.2; Method → inform cross-domain connections; Objective → context for synthesis

Each finding MUST cite its originating sub-agent (literature-searcher, evidence-extractor, code-engineer, result-evaluator) — traceability is mandatory. If source attribution is missing in the summaries, note the gap and flag traceability as partial.

Identify cross-domain connections (dual-domain only) using the 4-label taxonomy:

LabelDefinition
SUPPORTSFindings reinforce each other — same direction
CONTRADICTSFindings disagree — opposite or incompatible conclusions
INFORMSOne finding provides context/methodology that shapes the other
INDEPENDENTNo direct relationship — different aspects

Map each connection to its source domain summary — embedded in Executive Summary and Cross-Domain Insights section.

Surface contradictions verbatim — for every CONTRADICTS connection:

Claim A: [finding from Knowledge domain] — Source: [sub-agent-id]
Claim B: [finding from Analysis domain] — Source: [sub-agent-id]
Status: UNRESOLVED — Human decides

Never resolve, merge, or silently drop conflicting claims. The user is the sole arbiter.

Single-domain handling: Sections 4–5 note [SINGLE DOMAIN]. Intra-domain contradictions (if any) are surfaced labeled [INTRA-DOMAIN].

Phase 3 — Format and Write the Report

Determine output format:

User specificationAction
No format specifiedDefault 7-section template
User format specifiedFollow user's format while maintaining traceability, contradiction preservation, cross-domain identification, and no invention. Always include Verdict.

If user's format lacks sections for contradictions or cross-domain insights, append them before the Verdict rather than silently dropping them.

Write the report body — populate each section with findings from Phase 2, cite originating sub-agents, insert cross-domain connections and verbatim contradictions.

Verdict assignment:

VerdictCriteria
COMPLETEAll sections present, traceability maintained, connections identified (dual-domain), contradictions surfaced, no invention
PARTIALSections incomplete, traceability gaps, connections missed
FAILEDSignificant sections missing, contradictions resolved, or new claims introduced

Verdict section is always included regardless of format.

Phase 4 — Output Verification

Self-check before delivery:

  • Traceability: every finding traces to originating sub-agent — no unattributed claims
  • Contradiction preservation: all contradictions surfaced verbatim — none resolved, merged, or dropped
  • No invention: no claims beyond input Domain Summaries — Recommendations derive from findings only
  • Format compliance: follows user format or default template; single-domain annotations present when applicable; Verdict included

Step 3: Document Results

Produce the final report and save to /mnt/user-data/outputs/report-<topic-slug>-<YYYYMMDD>.md. Present file path back to the orchestrator.

Default Report Template

Used when user does not specify a format:

## 1. Executive Summary
[2-3 sentence overview of the research question, key findings, and overall verdict]

## 2. Knowledge Research Findings
### 2.1 Key Findings
- [finding] — Source: [sub-agent-id] — Confidence: [level]

### 2.2 Evidence Landscape
- [evidence summary with provenance status]

### 2.3 Identified Gaps
- [gaps in literature coverage]

## 3. Data Analysis Findings
### 3.1 Key Results
- [result] — Method: [method] — Evaluation: [PASS/FAIL/CONDITIONAL]

### 3.2 Robustness & Limitations
- [robustness assessment + limitations]

## 4. Cross-Domain Insights
- [connection: knowledge finding ↔ data finding] — Type: [SUPPORTS/CONTRADICTS/INFORMS/INDEPENDENT]
[SINGLE DOMAIN — no cross-domain analysis]

## 5. Contradictions (Unresolved)
- [Claim A] vs [Claim B] — Human decides.
[SINGLE DOMAIN — no inter-domain contradictions]

## 6. Limitations & Gaps
- [combined limitations from both domains]
- [coverage gaps not addressed]

## 7. Recommendations
- [actionable next steps with rationale]

### Verdict
- COMPLETE: all sections present, traceability maintained, cross-domain connections identified, contradictions surfaced, no content invention
- PARTIAL: some sections incomplete or traceability gaps or cross-domain connections missed
- FAILED: significant sections missing or contradictions silently resolved

Domain-Specific Report Patterns

DomainCommon report shapeTypical connection types
Biology — Mechanism + validationKnowledge: pathway mechanisms ↔ Analysis: experimental validationINFORMS, SUPPORTS
Biology — Biomarker + statisticsKnowledge: candidate biomarkers ↔ Analysis: sensitivity/specificitySUPPORTS, CONTRADICTS
Chemistry — Reaction methodology + yieldKnowledge: established procedures ↔ Analysis: yield optimization dataSUPPORTS, INFORMS
Materials — Structure-propertyKnowledge: known correlations ↔ Analysis: measured propertiesSUPPORTS, CONTRADICTS
Finance — Factor analysisKnowledge: academic factor models ↔ Analysis: backtest resultsSUPPORTS, CONTRADICTS

Boundary

Forbidden:

  • Do NOT add new findings, claims, or data — all content originates from Domain Summaries
  • Do NOT resolve contradictions between domains — surface them verbatim for the user
  • Do NOT evaluate result validity — evaluation belongs to result-evaluator
  • Do NOT perform analysis or research — those belong to their respective domain sub-agents
  • Do NOT silently drop or merge conflicting claims from different domains

Mandatory:

  • MUST produce output in user's specified format if provided; otherwise MUST use default 7-section template
  • MUST maintain traceability — every finding cites its originating sub-agent
  • MUST identify and label cross-domain connections explicitly (SUPPORTS/CONTRADICTS/INFORMS/INDEPENDENT) when both Domain Summaries are present
  • MUST surface contradictions verbatim — never resolve, merge, or silently drop conflicting claims
  • MUST NOT introduce any claims not present in the Domain Summaries
  • MUST note [SINGLE DOMAIN — no cross-domain analysis] in Section 4 and [SINGLE DOMAIN — no inter-domain contradictions] in Section 5 when only one domain is present
  • MUST map each cross-domain insight and each contradiction to the source Domain Summary — embedded in Executive Summary and Cross-Domain Insights

Error Handling

Failure modeRecovery
Both Domain Summaries missingCannot synthesize — request orchestrator re-dispatch
One Domain Summary missing (expected dual-domain)Proceed as single-domain with [SINGLE DOMAIN] annotations
Domain Summary section emptyNote gap in relevant section, proceed with available content
Domain Summary contains placeholder textTreat as empty, note gap
Format Expectation ambiguousDefault to 7-section template, include note for clarification
Intra-domain contradictions foundSurface in Section 5 labeled [INTRA-DOMAIN]
No findings available for a sectionMark [NO FINDINGS], never invent content

Complete Example

Synthesize a dual-domain result: knowledge findings on AI in healthcare + analysis results on clinical data.

  1. Read inputs — knowledge_summary.md and analysis_summary.md both present and substantive
  2. Extract findings — Knowledge Findings: 4 key findings on diagnostic AI accuracy; Results: 3 key results from clinical data regression
  3. Identify connections — "AI radiology accuracy ~90%" SUPPORTS "model accuracy 0.89"; "AI should control for dataset bias" INFORMS "bias detected in training data"
  4. Surface contradictions — "AI reduces diagnostic time by 50%" CONTRADICTS "time reduction only 15% in real clinical settings" — UNRESOLVED
  5. Write report — 7-section template, all findings with source attribution, Verdict COMPLETE
add-product-e2e-evalpaperclipai/paperclipAdd or extend a Paperclip full-stack runner E2E workflow, fixture, matcher, or report evidence path for local or Daytona execution.Stars: 115kUpdated Claude CodeCodexVerifiedInstall
add-runner-evalpaperclipai/paperclipAdd or extend a Paperclip Runner protocol evaluation definition, roster, assertion, or report fixture with provenance and narrow validation.Stars: 115kUpdated Claude CodeCodexVerifiedInstall
diagnose-why-work-stoppedpaperclipai/paperclipCompatibility entry point for failed, stalled, looping, or over-recovered Paperclip work. Use fix-failed-paperclip-run for the full diagnosis, regression, repair, real verification, and green-PR workflow; honor an explicitly requested diagnosis-only scope.Stars: 115kUpdated Claude CodeCodexVerifiedInstall
fix-failed-paperclip-runpaperclipai/paperclipDiagnose and fix failed, stuck, slow, or incorrectly blocked Paperclip tasks and runs. Use with a task/run URL or failure report to trace the root cause, prove a regression red then green, check second-order effects, verify the real workflow, and deliver a PR with passing checks.Stars: 115kUpdated Claude CodeCodexVerifiedInstall