Search-Synth: Multi-Agent Research with Fact-Checking

Purpose

Conduct thorough web research on a topic using multiple agents, cross-verify findings through fact-checking, and produce a synthesized "Best and Verified" response.

Required Inputs

Parameter	Description
topic	The research question or subject to investigate.
output_file	File path where the synthesized result will be written.
agents	Explicit list of participating agents, with each agent's search capability noted (see below).

Agent Roster and Search Capability

The invoker must declare which agents will participate and confirm which ones have web-search capability. Example:

agents:
  - name: CC
    role: MANAGER
    can_search: false
  - name: Codex
    role: SEARCHER
    can_search: true
    search_tool: <declared web-search tool>
  - name: Codex-peer-02
    role: SEARCHER
    can_search: true
    search_tool: <declared web-search tool>

Default roster: use Claude Code and Codex peers. Gemini CLI, Antigravity CLI, and Antigravity Desktop may participate only when the human explicitly selects them for this search run. Do not add them merely because they have a helper or search capability.

Pre-flight Check (Mandatory)

Before starting any phase, MANAGER must verify the roster:

1. At least two agents must have can_search: true.
2. At least one agent must be designated MANAGER (the MANAGER does not search).
3. A CHIEF_AUTHOR must be designated (defaults to MANAGER if not specified).

If the roster cannot satisfy search requirements -- for example, no agent has web search, or fewer than two searchers are available -- STOP immediately and report the mismatch. Do not improvise, simulate, or pretend to search.

Roles

Role	Responsibilities
MANAGER	Orchestrates phases, assigns tasks, does NOT search.
SEARCHER	Uses its declared search tool to research, writes expository response, fact-checks peers.
CHIEF_AUTHOR	Writes the final "Best and Verified" synthesis. Defaults to MANAGER; can be reassigned.

Protocol

Phase 1: Search and Respond

MANAGER assigns the topic to all SEARCHERS:

[AGENT] - MANAGER here. TASK: Research "[TOPIC]"

1. Use YOUR search tool to research this topic thoroughly
2. Write a NARRATIVE response (3-5 paragraphs) that:
   - Provides a well-considered, useful, actionable answer
   - Synthesizes findings into coherent prose (NOT just a list of results)
   - Uses proper citations [1] for non-common facts
3. Add your response to [OUTPUT_FILE] under "## [AGENT] Response"
4. BELOW your response, add "### [AGENT] Search Terms & Raw Results" with:
   - Your actual search query
   - URL + Title/Blurb table for each source
5. Signal [PHASE1_DONE] when complete

IMPORTANT: Response must be expository/narrative FIRST, then raw results.
Do NOT provide "next 10" style listings. Synthesize and explain.

All SEARCHERS work in parallel.

Phase 2: Fact-Check

After all SEARCHERS signal [PHASE1_DONE], MANAGER assigns fact-checking. Each SEARCHER fact-checks every other SEARCHER's response:

[AGENT] - MANAGER here. TASK: Fact-check the other responses.
1. Read the responses from [OTHER_AGENTS]
2. For EACH assertion/claim, verify via web search
3. Add your fact-checks to [OUTPUT_FILE] under "## Fact Checks" -> "### [AGENT] Fact-Check of [OTHER]"
4. Format: Claim -> Verdict (VERIFIED / FALSE / UNVERIFIABLE) -> Source
5. Signal [PHASE2_DONE] when complete

Phase 3: Best and Verified Synthesis

After all fact-checks complete, MANAGER prompts CHIEF_AUTHOR:

CHIEF_AUTHOR - MANAGER here. TASK: Write the "Best and Verified" response.
1. Read ALL responses from Phase 1
2. Read ALL fact-checks from Phase 2
3. Write a comprehensive response that:
   - Captures ALL good/accurate content from each original response
   - EXCLUDES any assertions marked FALSE by fact-checkers
   - Synthesizes into a coherent, authoritative answer
4. Add an "### Unverified Claims" section listing assertions that:
   - May be true but could not be verified from authoritative sources
   - Are widely reported but not officially documented
5. Add to [OUTPUT_FILE] under "## Best and Verified Response"
6. Signal [DONE] when complete

Output File Structure

# [TOPIC] Research Synthesis

**Skill:** search-synth
**Date:** [DATE]
**Searchers:** [list of SEARCHER agents]
**Chief Author:** [CHIEF_AUTHOR agent]

---

## Phase 1: Initial Responses

### [Agent A] Response
[expository response]

#### [Agent A] Search Terms & Raw Results
**Query:** "[search terms]"
| URL | Title/Blurb |
|-----|-------------|
| ... | ... |

### [Agent B] Response
...

---

## Phase 2: Fact Checks

### [Agent A] Fact-Check of [Agent B]
| Claim | Verdict | Source |
|-------|---------|--------|
| ... | VERIFIED/FALSE/UNVERIFIABLE | ... |

### [Agent B] Fact-Check of [Agent A]
...

---

## Phase 3: Best and Verified Response

**Chief Author:** [CHIEF_AUTHOR]

[Final synthesized response incorporating all verified information]

### Unverified Claims

The following assertions appear plausible but could not be verified
from authoritative sources:

- [Claim 1] -- Reported by [source], not officially documented
- ...

---

**[DONE]**

Termination Signals

Signal	Meaning
[PHASE1_DONE]	Searcher completed response + raw results
[PHASE2_DONE]	Searcher completed fact-checks of others
[DONE]	Chief author completed final synthesis
[BLOCK]	Cannot proceed (missing input, agent unresponsive)
[MISMATCH]	Roster does not satisfy search requirements; skill halted

Error Handling

Search Capability Mismatch

If MANAGER determines that the roster lacks sufficient search capability (fewer than two agents with can_search: true), the skill must:

1. Signal [MISMATCH].
2. Report which agents were provided and what capabilities they declared.
3. Halt. Do not attempt to proceed with insufficient search coverage.

Timeout with POKE

Instead of immediately abandoning slow agents, MANAGER sends a POKE:

At 60s (first timeout):

[AGENT] - MANAGER here. POKE: You haven't reported Phase [N] results.
Please complete your search and post findings now.
If blocked, signal [BLOCK] with reason.

At 120s (final timeout):

• If still no response after POKE: MANAGER signals [BLOCK] and proceeds without that agent.
• Log: "[AGENT] unresponsive after POKE, proceeding with available agents"

Other Error Conditions

• If fewer than 2 searchers complete Phase 1: skill exits with [INCOMPLETE].
• If CHIEF_AUTHOR unavailable: MANAGER may reassign to next available agent.

Citation Requirements

All agents must:

• Use proper citations [1] for non-common facts derived from web sources.
• Avoid raw markdown links in narrative text; use citation format instead.
• Include full URLs only in the "Raw Results" table section.

Example Prompts

Basic

Use search-synth.
Topic: "What is retrieval-augmented generation (RAG)?"
Output: projects/research/rag-synthesis.md
Agents:
  - CC (MANAGER, no search)
  - Codex (SEARCHER, declared web-search tool)
  - Codex-peer-02 (SEARCHER, declared web-search tool)
Chief author: CC

Two searchers with explicit capability

Use search-synth.
Topic: "History of computational law"
Output: projects/research/complaw-history.md
Agents:
  - CC (MANAGER + CHIEF_AUTHOR, no search)
  - Codex (SEARCHER, declared web-search tool)
  - CC-instance-2 (SEARCHER, brave_web_search via MCP)

Notes

• No scaffolding code -- this skill relies on well-written instructions, not automation.
• Manager is orchestrator only -- does not participate in search or fact-checking.
• Parallel execution -- Phase 1 searchers work simultaneously.
• Sequential phases -- Phase 2 starts only after all Phase 1 signals are received.
• Explicit capability -- never assume an agent can search; the roster must declare it.