Claude Skills by Snowflake-AI-Research

Ablation variant of bbs-coordination-web with the idle peer-review protocol ("When Idle") removed (GATE 2 OFF — no board review). Posting and reading of findings is retained. Used by the R1 and R0 arms of the review-gate ablation; do not use for production runs. Load this skill before using the post_to_bbs tool.

Shared Bulletin Board System (BBS) coordination protocol for web-research swarm subagents. Covers posting and reading from BBS, channel descriptions, structured data requirements, and idle review behavior. Load this skill before using the post_to_bbs tool.

Workflow for writing and executing code (bash, Python) to accomplish computation, data processing, and scripting tasks. Covers planning, dependency management, error handling, file organization, and result verification.

Deep chain-of-thought analysis workflow using the reasoning tool. Covers when to use reasoning, how to provide context, verifying findings, and producing well-supported conclusions.

Direct messaging protocol for targeted agent-to-agent communication. Covers sending and receiving private messages, when to use DM vs BBS, and handling incoming messages during idle periods. Load this skill before using the send_message tool.

DM coordination protocol optimized for 2-agent duo teams. The main worker drives the analysis while the auditor reviews, challenges, and fills gaps. With only one recipient per DM, sharing is cheap. Load this skill to understand the duo communication model.

Workflow for analyzing ambiguous data questions. Covers decomposing questions, identifying implicit assumptions and edge cases, verifying interpretations with SQL evidence, and guiding the team toward the correct interpretation.

Event-driven orchestrator workflow for coordinating a team of subagents via task board delegation in DM-only mode (no BBS) with agent_spawnable support (create_agent, assign_to). Unlike the blocking variant, the orchestrator ends its turn after creating tasks and gets woken up by subagent DM notifications between turns. Covers the workflow (understand, spawn team, delegate, end turn, review DMs, mediate, report), visualization guidelines, citation protocol, and task creation best practices.

Event-driven orchestrator workflow for coordinating a team of subagents via task board delegation in DM-only mode (no BBS). Unlike the blocking variant, the orchestrator ends its turn after creating tasks and gets woken up by subagent DM notifications between turns. Covers the workflow (understand, delegate, end turn, review DMs, mediate, report), visualization guidelines, citation protocol, and task creation best practices.

Orchestrator workflow for coordinating a team of subagents via task board delegation in DM-only mode (no BBS) with agent_spawnable support (create_agent, assign_to). Covers the workflow (understand, spawn team, delegate, wait, review, report), visualization guidelines, citation protocol, and task creation best practices.

Orchestrator workflow for coordinating a team of subagents via task board delegation in DM-only mode (no BBS). Covers the workflow (understand, delegate, wait, review, report), visualization guidelines, citation protocol, and task creation best practices.

Orchestrator workflow for coordinating on-demand subagents via task board delegation in DM-only mode (no BBS) with dynamic scaling.

Ablation variant of swarm-orchestration-dynamic-web with BOTH the board-review narration (auditor auto-spawn + idle review, GATE 2 OFF) and the final-verification / alternative-forcing instructions (GATE 3 OFF) removed. Used by the R1 and R0 arms of the review-gate ablation; do not use for production runs.

Ablation variant of swarm-orchestration-dynamic-web with the orchestrator's final-verification / alternative-forcing instructions removed (GATE 3 OFF). Board review (auditor + idle review) is retained. Used by the R2 arm of the review-gate ablation; do not use for production runs.

Orchestrator workflow for coordinating on-demand web-research subagents via task board delegation with dynamic scaling.

Orchestrator workflow for coordinating on-demand subagents via task board delegation with dynamic scaling. Workers are spawned when tasks are created. Covers the workflow, assign_to for context reuse, visualization guidelines, citation protocol, and task creation best practices.

Orchestrator workflow for coordinating a team of web-research subagents via task board delegation. Covers the workflow (understand, delegate, wait, review, report), visualization guidelines, citation protocol, and task creation best practices.

Orchestrator workflow for coordinating a team of subagents via task board delegation. Covers the workflow (understand, delegate, wait, review, report), visualization guidelines, citation protocol, task creation best practices, and BBS channel reference.

Task lifecycle protocol for web-research subagents in DM/Duo mode (no BBS). Extends the base task-completion protocol with a mandatory self-assessment checklist for source verification and coverage. Results are shared through `complete_task` summaries and targeted `send_message` DMs — there is no Bulletin Board System.

Task lifecycle protocol for web-research swarm subagents. Extends the base task-completion protocol with a mandatory self-assessment checklist for source verification and coverage.

Shared task lifecycle protocol for swarm subagents. Covers executing assigned tasks, posting results, and completing tasks with summaries.

Tool usage guidelines for corpus research agents running in DM/Duo mode. Covers parallel tool calling, the corpus search-then-fetch workflow, and sharing results via direct messaging and task-completion summaries.

Tool usage guidelines for corpus research agents. Covers parallel tool calling, search strategy, and the search-then-fetch workflow for the static document corpus.

Tool usage guidelines for web research agents in DM/Duo mode (no BBS). Covers tool reference, search-first principle, and the principle of verifying before claiming. Sharing happens via direct messaging and task-completion summaries — there is no Bulletin Board System.

Tool usage guidelines for web research agents. Covers tool reference, search-first principle, and the principle of verifying before claiming.

Tool usage guidelines for code execution tasks. Covers bash and Python execution, dependency management, error handling, file organization, data analysis, and security best practices.

Image understanding guidance for agents with vision enabled. Instructs agents to use read_file for viewing images directly instead of writing OpenCV/PIL code to parse them programmatically.

Ablation variant of web-research-corpus with the iterative Search->Assess->Decide reflection loop and its strict completion criteria removed (GATE 1 OFF — no local reflection). The agent performs a single thorough search pass. Used by the R0 arm of the review-gate ablation; do not use for production runs.

Web research methodology adapted for static document corpus retrieval. Covers iterative search loop with corpus search and full-document fetch, strict completion criteria, information redundancy principle, and automatic source quality scoring.

Web research methodology with iterative search loop, strict completion criteria, information redundancy principle, and automatic source quality scoring. Covers search strategy, source evaluation, cross-verification, and handling conflicting information.