ワンクリックで
search-synth
Multi-agent web research with cross-fact-checking and verified synthesis.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
メニュー
Multi-agent web research with cross-fact-checking and verified synthesis.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | search-synth |
| description | Multi-agent web research with cross-fact-checking and verified synthesis. |
| metadata | {"owner":"interlateral","version":"0.1"} |
Conduct thorough web research on a topic using multiple agents, cross-verify findings through fact-checking, and produce a synthesized "Best and Verified" response.
| 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). |
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.
Before starting any phase, MANAGER must verify the roster:
can_search: true.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.
| 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. |
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.
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
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
# [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]**
| 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 |
If MANAGER determines that the roster lacks sufficient search capability (fewer than two agents with can_search: true), the skill must:
[MISMATCH].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):
[BLOCK] and proceeds without that agent.[INCOMPLETE].All agents must:
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
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)
Coordination guidance for the narrow case of multiple desktop-backed agents active concurrently in the same desktop-backed coordination context. Use ONLY when two or more desktop peers are running at the same time; normal CLI agents and single desktop peers should ignore this skill. Read this BEFORE concurrent desktop-to-desktop handoffs to keep sender labels unique, identity routing cooperative, and material handoffs proven.
Mandatory near-live communication discipline ("the Law") for the two DESKTOP agents (claude-desktop and codex-desktop) on the Interlateral mesh. Read this BEFORE any desktop-to-desktop peer communication, peer-superset, ready-rock-quartet, sprint-overseer, gate-keeper, sprint handoff, or substantial delegation. It forces correct live comms and prevents the recurring failure modes. Normal CLI agents, Warp/terminal peers, and single non-desktop workflows should ignore this skill and use the ordinary repo protocols.
Join the Antigravity CLI (agy, Gemini 3.5 Flash) to the Interlateral mesh as a native CLI peer with its own tmux session, identity stamping, and the agy.js send helper. Use this instead of the Antigravity desktop-app CDP path.
Run a two-agent proxy approval council for autonomous sprint gates. Use when a sprint needs delegated approval for human gates such as deploy, live mutation, push, cleanup, skip, or shortened watch decisions, while requiring two independent agents to reach consensus that the human principal would approve under written policy and preferences.
Initialize the standard Interlateral two-agent CLI mesh from a bootstrap agent by running me.sh, launching Claude Opus 4.7 plus Codex 5.5, proving live ACK, and waiting for "Reporting for Duty!". The bootstrap agent is not part of the mesh.
Operate or troubleshoot Interlateral CLI mesh transport: tmux socket discovery, direct-send helpers, comms.md ledger, identity stamping, idle checks, ACK proof, and safe prompt injection. This is comms mechanics only, not a collaboration workflow.