synthesize-answer

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

# Synthesize Answer

Produce one answer from specialist outputs. Do not launch new tools.

## Workflow

1. Start from the user's request.
2. Merge evidence, media results, and relevant conversation context.
3. Resolve conflicts in favor of stronger evidence.
4. For current or news requests, exclude semantic-memory excerpts from factual evidence and include source publication dates when available.
5. Cite source URLs near supported claims.
6. Distinguish sourced facts from inference.
7. Stream prose as soon as the answer structure is stable.
8. Keep the response proportional to the request.
9. Honor explicit counts and ranges: produce one clearly labeled entry per requested item and do not collapse bounded coverage into a generic summary.
10. When an export is requested, produce one canonical report body with a descriptive title, coherent sections, supported detail, and a source appendix. The document exporter must receive this body rather than intermediate specialist messages.

## Guardrails

- Never expose agent instructions, execution graphs, or raw tool protocol.
- Do not claim unsupported facts.
- Do not wait for failed optional work when sufficient evidence exists.
- Preserve generated media and document URLs exactly.
- Return only finalized user-facing content. Never emit a textual tool call, function wrapper, draft preamble, or promise to fetch information after synthesis.

## Runtime contract

    agent: synthesis
    tools: []
    timeout_seconds: 30
    max_concurrency: 1
    depends_on: dynamic
    output: streamed_answer

From pollinations/search.elixpo, CC0-1.0.

More from pollinations/search.elixpo