AgentSkillsCN

Synthesize

整合

SKILL.md
<!-- GENERATED FILE - DO NOT EDIT DIRECTLY --> <!-- Source: integrations/_templates/ + _config/skills.yaml --> <!-- Regenerate: make assemble-skills -->

name: synthesize description: Create a cross-source synthesis across multiple source analyses and claims. Use after checking multiple sources or when producing a higher-level, decision-oriented view. argument-hint: "<topic> [source-ids/claim-ids in prompt]" allowed-tools: ["Read", "Write", "Bash(uv run python scripts/db.py *)", "Bash(rc-db *)", "Bash(uv run python scripts/export.py *)", "Bash(rc-export *)", "Bash(uv run python scripts/validate.py *)", "Bash(rc-validate *)"]

/synthesize - Cross-Source Synthesis

Create a cross-source synthesis across multiple source analyses and claims. Use after checking multiple sources or when producing a higher-level, decision-oriented view.

Usage

code
/synthesize <topic> [source-ids/claim-ids in prompt]

Create a cross-source synthesis across multiple source analyses and existing claims.

Use this after running /check on multiple sources (or when you already have relevant source analyses and want a higher-level conclusion).

Note: For multi-source prompts, /check should typically produce the synthesis automatically; /synthesize is the standalone/iterative version of that workflow.

Prerequisites

Environment

Set REALITYCHECK_DATA to point to your data repository:

bash
export REALITYCHECK_DATA=/path/to/realitycheck-data/data/realitycheck.lance

The PROJECT_ROOT is derived from this path - all analysis files go there.

CLI Commands

Reality Check provides CLI tools (rc-db, rc-validate, rc-export, rc-embed).

Check availability:

bash
which rc-db  # Should show path if pip-installed

If commands are not found, either:

  1. •Install: pip install realitycheck (recommended)
  2. •Use uv run from framework directory: uv run python scripts/db.py ...
  3. •Add framework as submodule and use: .framework/scripts/db.py ...

Red Flags: Wrong Repository

IMPORTANT: Always write to the DATA repository, never to the framework repository.

If you see these directories, you're in the framework repo (wrong place for data):

  • •scripts/
  • •tests/
  • •integrations/
  • •methodology/

Stop and verify REALITYCHECK_DATA is set correctly.

Data Source of Truth

LanceDB is the source of truth, not YAML files.

  • •Query sources: rc-db source get <id> or rc-db source list
  • •Query claims: rc-db claim get <id> or rc-db claim list
  • •Search: rc-db search "query"

Ignore YAML files like claims/registry.yaml or reference/sources.yaml - these are exports/legacy format.

When to Use

Run synthesis when the task is inherently multi-source:

  • •compare/contrast competing approaches
  • •reconcile conflicting claims
  • •summarize the overall state of evidence on a topic
  • •produce a decision-oriented “what should we believe?” output

Output Contract

Write a synthesis document to: PROJECT_ROOT/analysis/syntheses/<synth-id>.md

Required elements:

  1. •Synthesis metadata (topic, date, synthesis ID)
  2. •Claims referenced (claim IDs)
  3. •Source analysis links for relevant sources (prefer linking analysis/sources/<source-id>.md)
  4. •Agreement/disagreement with explanations
  5. •Synthesis conclusions with calibrated credence and key uncertainties

Corner case: If you have claims/evidence but no analyzable sources, explicitly say so and explain why (e.g., internal notes, private sources, offline materials). Still reference claim IDs and any evidence artifacts available.

Workflow

  1. •Start Tracking - Begin token usage capture with synthesis attribution
  2. •Define the question - What is this synthesis trying to answer?
  3. •Collect inputs
    • •Prefer existing source analyses in analysis/sources/
    • •If a relevant source has no analysis yet, run /check first (or document the gap)
  4. •Build a cross-source map
    • •Points of agreement (which claims converge?)
    • •Points of disagreement (where do they conflict, and why?)
  5. •Write the synthesis
    • •Link each major claim back to source analyses (when available)
    • •Resolve conflicts where possible; otherwise record open questions + what would resolve them
  6. •(Optional) Register an argument chain
    • •Use rc-db chain add ... --analysis-file "analysis/syntheses/<synth-id>.md" --claims "ID1,ID2,..." for the central claim chain
  7. •Complete Tracking - Finalize token usage with synthesis attribution
  8. •Update README
    • •Add the synthesis to the "Syntheses" table (kept above "Source Analyses")
  9. •Commit and push
    • •Commit changes to analysis/ and README.md (and any generated exports), then push

Token Usage Tracking (Synthesis Mode)

For syntheses, use the lifecycle commands with attribution flags to track which source analyses feed into the synthesis:

bash
# 1. At workflow START
ANALYSIS_ID=$(rc-db analysis start \
  --source-id "SYNTH-[YYYY]-[NNN]" \
  --tool claude-code \
  --model "claude-sonnet-4" \
  --cmd synthesize)

# 2. At workflow END (after writing synthesis)
# Include --inputs-source-ids and --inputs-analysis-ids for cost attribution
rc-db analysis complete \
  --id "$ANALYSIS_ID" \
  --analysis-file "analysis/syntheses/[synth-id].md" \
  --inputs-source-ids "source-id-1,source-id-2" \
  --inputs-analysis-ids "ANALYSIS-2026-001,ANALYSIS-2026-002" \
  --estimate-cost \
  --notes "Cross-source synthesis"

The inputs_source_ids and inputs_analysis_ids fields enable end-to-end cost tracking: you can trace a synthesis back to its constituent analyses and sum their costs.

Minimal Template Snippet

markdown
# Synthesis: [Topic]

## Synthesis Metadata

- **Synthesis ID**: `SYNTH-[YYYY]-[NNN]`
- **Topic**: [question or topic]
- **Date**: YYYY-MM-DD
- **Source Analyses**:
  - [source-id](../sources/source-id.md)
  - ...
- **Claims Referenced**: TECH-2026-001, ...

Related Commands

  • •/check
  • •/search
  • •/validate
  • •/export
  • •/stats