questionably-ultrathink-skill

What it does

The skill guides an agent to handle complex problems by creating a structured, multi-step reasoning pipeline. It separates question generation, claim formation, and verification into isolated roles, then combines verified results into a final answer. It uses a graph-based orchestration to manage multiple atomic questions, with explicit steps to generate claims, spawn verifiers for each claim, and a maintainer to produce the final answer.

How it works

• It defines an architecture where an orchestrator delegates tasks to several agents in isolation: a Graph Generator builds a DAG of questions; a Claim Generator creates claims and pre-creates verifier files containing only the verification questions; Verifiers (one per verification question) research and answer in complete isolation without seeing the claim text; a Verification Maintainer cross-checks claims against verifier answers and synthesizes the final answer; a Graph Maintainer could contract solved atoms with dependencies.
• Per atom, the workflow proceeds: read question.md, generate claims.md and verifiers/.md; verifiers are pre-created with only the verification questions; verifiers are executed to produce answers; the maintainer reads claims.md and verifiers/.md to create answer.md.
• Background execution is encouraged to run verifications in parallel and improve throughput; you can spawn verifiers in background and later check for completion via TaskOutput or by grepping for "# Answer" in verifier files.
• The process is stateful only by file existence and content: question.md → atom created; claims.md → claims generated; answer.md → verification complete.

When to use it

• Activation triggers include phrases like "be thorough", "analyze carefully", or "make sure this is right".
• Use for Complex multi-part questions, Architecture or security decisions, high-stakes technical decisions, or debugging complex issues.
• Engage when you want rigorous, factored verification with isolation between claim development, verification, and final synthesis.

What it can touch

• It uses Task tool calls with subagent_type values:
  • "questionably-ultrathink:aot-graph-generator" to create the DAG
  • "questionably-ultrathink:cove-claim-qs" to generate claims.md and pre-create verifiers (verifiers/{N}.md)
  • "questionably-ultrathink:cove-verifier" to populate each verifier file with answer data
  • "questionably-ultrathink:cove-verification-maintainer" to create final answer.md from claims and verifiers
  • Optional: background execution flags like run_in_background: true

Caveats

• License: MIT
• The verifier reads ONLY its pre-created file containing the verification question and has ZERO knowledge of the claim text.
• Verifiers are expected to overwrite their verifier file with question + answer + confidence + sources.
• The final answer is synthesized by a separate maintainer, which reads claims.md and verifiers/*.md to produce answer.md.