multi-source-synthesis

Multi-Source Synthesis

Combine multiple summaries into a single coherent synthesis with proper attribution, deduplication, and conflict handling.

When to Use This Skill

The model MUST activate this skill when:

• User requests: "combine these summaries", "synthesize results", "merge findings"
• Multiple sources have been individually summarized and need integration
• User asks to "pull together" or "integrate" information from multiple sources
• This is the "reduce" step after the "map" step (individual summarization)

Prerequisites

Before applying this skill, the model MUST ensure:

1. Each source has been individually summarized using the summarizer skill
2. Each individual summary followed Rule 1 (Read Before Summarizing) and Rule 2 (Extract Before Abstracting) from Fidelity Rules
3. Each summary follows the structured format from Structured Summary
4. Source attribution is present in each individual summary

Synthesis Workflow

The model MUST follow these steps in order:

Step 1: Deduplicate

Merge identical or overlapping information from different sources:

Rules:

• If the same information appears in multiple sources, merge into a single statement and cite ALL sources
• Use the most complete version of the information among the sources
• Do NOT drop any source citations during merging
• If two sources say essentially the same thing with different wording, use the clearer or more precise wording
• Preserve exact numbers and specifics (see Fidelity Rules Rule 3)

Example:

BEFORE DEDUPLICATION:
- Source A: "API rate limit is 100 requests per minute"
- Source B: "Rate limiting enforced at 100 req/min per key"

AFTER DEDUPLICATION:
- API rate limit is 100 requests per minute per key (sources: A, B)

Deduplication Decision Matrix:

Situation	Action
Same info, different sources	Merge, cite all sources, use most complete version
Same topic, different conclusions	Keep both, surface conflict explicitly
Different aspects of same topic	Keep separate, group under same theme cluster
Different people/viewpoints	Keep separate, attribute to each source

Step 2: Cluster by Theme

Group related findings by topic or theme, NOT by source:

Rules:

• Organize information by conceptual theme, not by which source it came from
• Each cluster should answer a specific question or cover a specific subtopic
• Findings from multiple sources should be integrated within each cluster
• Avoid patterns like "Source A says..., Source B says..." unless explicitly comparing viewpoints

Example structure:

## Authentication Methods

The API supports JWT authentication (sources: A, C) and OAuth 2.0 (source: B). JWT tokens expire after 24 hours (source: A), while OAuth tokens are configurable per client (source: B).

Step 3: Rank Clusters

Order clusters by relevance to the user's query or task:

Rules:

• Lead with the most important information (BLUF - Bottom Line Up Front)
• If user query is specific, rank clusters by how directly they answer the query
• If user query is broad, rank by significance or impact
• Consider dependencies: foundational concepts before advanced details

Step 4: Assess Confidence Across Sources

Calculate confidence for each synthesized claim based on source agreement:

Confidence Scoring Rules:

Source Profile	Confidence Level
Multiple fresh authoritative sources agree	High
Single authoritative source, recent	Medium-High
Multiple sources but somewhat dated	Medium
Single source or informal source	Medium-Low
Old data, conflicting sources, or ambiguous	Low

The model MUST:

• Include confidence level in the synthesis frontmatter
• Add confidence notes explaining the assessment
• Surface low-confidence claims explicitly in the Uncertain section

Confidence calculation factors (from Fidelity Rules Rule 6):

• Freshness: How recent is the source? Dated information reduces confidence
• Authority: Official docs > community posts > informal sources
• Agreement: Multiple sources agreeing increases confidence
• Completeness: Partial or truncated sources reduce confidence

Step 5: Synthesize

Produce a narrative synthesis with proper attribution:

Rules:

• Use the structured output format from Structured Summary
• Set source_type: multi-source in frontmatter
• List all source paths in source_path as a YAML list
• Attribute claims to their sources inline: (source: A) or (sources: A, B, C)
• Do NOT add information not present in any source
• Do NOT resolve conflicts by picking one version without stating the choice

Attribution format:

JWT tokens expire after 24 hours (source: API Documentation v2.3). However, the Developer Guide states tokens expire after 12 hours. This discrepancy requires clarification.

Step 6: Format Based on Result Count

Choose detail level based on the number of distinct findings after deduplication:

Result Count	Strategy	Format
Small (1-5 findings)	Present each finding with full detail	Bulleted list with inline attribution
Medium (6-15 findings)	Group by theme, provide detail within each theme	Themed sections with subsections
Large (16+ findings)	High-level summary + offer drill-down	Executive summary + detailed sections

Small result set example:

## Summary

Four key capabilities were identified:

- JWT authentication with 24-hour token expiration (source: API Docs)
- Rate limiting at 100 requests/min per API key (sources: API Docs, Developer Guide)
- WebSocket support for real-time updates (source: Developer Guide)
- Webhook notifications for asynchronous events (source: Integration Guide)

Large result set example:

## Summary

The synthesis covers 23 distinct API capabilities across 5 categories: authentication, rate limiting, real-time features, data formats, and error handling.

### Authentication (5 findings)

[Detailed subsection with attributions...]

### Rate Limiting (4 findings)

[Detailed subsection with attributions...]

[Additional sections...]

See "What Was Found" section for complete enumeration with source references.

Conflict Handling

When sources disagree, the model MUST surface conflicts explicitly:

Rules:

• NEVER silently pick one version when sources conflict
• State both versions with their sources
• Add the conflict to the Uncertain section
• If one source is clearly more authoritative or recent, note that but present both
• Recommend verification or clarification to the user

Conflict pattern:

## Uncertain

- Token expiration time: API Documentation v2.3 states 24 hours, but Developer Guide v1.8 states 12 hours. The API Documentation is more recent (2026-01-15 vs 2025-11-03), but both are official sources. Recommend testing or contacting support for clarification.

Multi-Source Output Format

Use the structured format from Structured Summary with these multi-source specifics:

YAML frontmatter:

---
source_type: multi-source
source_path:
  - /path/to/source-a.md
  - https://example.com/source-b
  - /path/to/source-c.pdf
summarized_at: "2026-02-06T14:32:00Z"
method: hybrid
word_count_source: null
word_count_summary: 847
compression_ratio: null
confidence: medium
confidence_notes: "Three sources from 2025-2026, all authoritative. One minor conflict on token expiration requires verification."
---

Body sections (all required):

1. Summary - The synthesized narrative
2. What Was Found - Deduplicated findings with multi-source attributions
3. What Was NOT Found - Items searched across all sources but absent
4. Uncertain - Conflicts, ambiguities, items requiring clarification
5. Sources - Complete list with access dates

Team Coordination Mode

When this skill is invoked after a team-based summarization (teammates instead of subagents), the synthesis workflow is the same but the input format differs:

Input: Teammate Messages vs Subagent Returns

Orchestration	Input to Synthesis
Subagents	Summary text returned directly from Agent tool calls
Teammates	Findings delivered via Teammate inbox messages to the leader

Cross-Checking During Summarization

In team mode, teammates message each other directly when they discover overlapping or contradictory information. This means some deduplication and conflict detection happens during the map step rather than only in this reduce step.

The leader MUST still run the full synthesis workflow (Steps 1-6) on the collected findings. Teammate cross-checking reduces but does not eliminate the need for post-hoc deduplication and conflict handling.

Leader Responsibilities

1. Collect all teammate findings from the inbox
2. Run the synthesis workflow (Steps 1-6) on the collected findings
3. Request shutdown for all teammates after synthesis is complete
4. Cleanup the team

SOURCE: Team coordination pattern from orchestrating-swarms skill (accessed 2026-02-06). Swarm pattern (self-organizing workers with shared task list) applied to multi-source summarization.

Output Rendering

1. Read template - Load the template file at ../summarizer/templates/{format_id}.md (default: structured). The template defines the schema, required sections, and fidelity constraints for the selected format.
2. Render - Produce output following the template's Schema section. Use the template's Example as a reference for structure and style.
3. Verify fidelity - Confirm the output satisfies the template's Fidelity Constraints and all applicable Fidelity Rules.

Anti-Patterns

The model MUST NOT:

• List results source-by-source instead of clustering by theme
• Bury the answer under methodology descriptions
• Present uncertain information with the same confidence as well-supported facts
• Silently resolve conflicts by picking one version
• Drop source citations during deduplication
• Re-summarize the individual summaries (violates Fidelity Rules Rule 5)
• Upgrade "not found in any source" to "doesn't exist"
• Omit the confidence assessment
• Fail to distinguish between single-source and multi-source agreement

Sources

This skill is adapted from Anthropic's knowledge-synthesis skill in the knowledge-work-plugins repository (accessed 2026-02-06). Synthesis workflow steps, confidence scoring methodology, and result set size strategies are derived from that source.