research-synthesis

name

research-synthesis

description

Synthesize evidence from multiple heterogeneous sources into a structured insight brief with confidence, contradictions, convergence/divergence, and decision implications. Use when the user asks to reconcile conflicting feedback, make sense of interviews, notes, transcripts, support signals, chats, research documents, or mixed evidence, identify what the data really says, explain underlying needs behind surface requests, or produce weighted research synthesis from multiple inputs.

research-synthesis

Synthesize multiple evidence sources into a decision-ready brief.

Goal

Turn messy multi-source input into:

• ranked insights
• confidence levels
• contradictions and segment splits
• explicit gaps / unknowns
• concrete implications

Core principle

Do not merely summarize inputs. Reconcile them.

The skill must distinguish between:

• what people said
• what likely happened
• what pattern is strong enough to count as an insight
• what remains only a hypothesis

Prefer a smaller number of defensible insights over a long list of weak observations.

Standard workflow

1. Frame the synthesis task

Capture:

• decision to support
• topic / scope
• time window
• source types available
• obvious gaps in evidence

If the user did not specify the decision context, infer it cautiously and say so.

2. Inventory sources and provenance

List each source class before interpreting it. For each source, note when possible:

• source type
• count / sample size
• date range
• direct vs second-hand
• likely bias or blind spots

3. Normalize signals

Convert messy evidence into comparable units.

Minimum normalization fields:

• source
• signal
• tentative underlying need / issue
• affected segment
• evidence strength
• notable quote or example if needed

4. Weight evidence

Apply weighting heuristics from references/weighting-model.md.

Rules:

• weighting guides judgment; it does not replace it
• repeated weak signals do not automatically beat fewer high-quality signals
• one vivid anecdote must not dominate the synthesis

5. Detect convergence

Look for the same underlying need across multiple independent sources.

Treat convergence as stronger when:

• it appears in 3+ independent signals
• it spans different source types
• it recurs across roles, contexts, or time windows

6. Detect divergence and segmentation

Do not collapse genuine disagreement into a fake consensus.

When signals conflict, test whether the disagreement is explained by:

• different user segments
• different contexts of use
• different experience levels
• different task frequencies
• artifact bias in the source

7. Separate surface requests from underlying needs

Translate feature asks and complaints into deeper needs where justified.

Example pattern:

• surface request: "make export faster"
• possible underlying need: "users do not trust that output will be ready when they need it"

Do not over-interpret. If the inference is weak, mark it as a hypothesis.

8. Produce the synthesis

Output a compact brief with:

• strongest insights first
• confidence for each insight
• evidence and counter-evidence
• unresolved contradictions
• explicit unknowns
• practical implications / next moves

Output contract

Minimum output sections:

• Sources included
• Signal quality / confidence note
• Top insights
• Contradictions / divergent signals
• What the data does not tell us
• Recommended next moves

Use references/output-patterns.md for exact structure.

Quality gates

Before finalizing, verify:

• the synthesis is not just a paraphrase of notes
• each major insight has visible evidence behind it
• conflicting evidence is not hidden
• confidence is earned, not decorative
• gaps and unknowns are stated explicitly
• recommendations follow from evidence, not from preference

Mode guidance

Interview-heavy mode

Use when the corpus is mostly interviews, call transcripts, or qualitative sessions. Read references/source-types.md and emphasize recurring themes over memorable quotes.

Mixed-evidence mode

Use when evidence comes from multiple formats: chats, docs, transcripts, graph cards, support signals, analytics summaries. Normalize first, then synthesize.

Conflict-reconciliation mode

Use when the user explicitly says the sources disagree. Prioritize:

• surfacing contradictions clearly
• testing segmentation hypotheses
• separating strong conclusions from unresolved disputes

What not to do

• Do not treat loud signals as representative by default.
• Do not flatten all sources into equal credibility.
• Do not present hypotheses as conclusions.
• Do not hide uncertainty to make the output look cleaner.
• Do not generate more insights than the evidence can support.

Read references as needed

• references/weighting-model.md — source weighting and confidence rules
• references/output-patterns.md — canonical synthesis output patterns
• references/source-types.md — handling interviews, notes, transcripts, graph context, chats, and mixed evidence