Claude Code’s 16-Step Research Assembly Line

What it does

Hyperresearch is a Claude Code skill that orchestrates a 16-step research pipeline across a roster of specialized subagents. Depending on the tier you request, it spends 30 minutes to over two hours decomposing queries, sweeping sources, investigating depth loci, drafting from multiple angles, and adversarially patching the result. Every fetched source is stored as plain markdown with YAML frontmatter in a local vault backed by a rebuildable SQLite index, so your research corpus compounds across sessions.

The interesting bit

The project treats the canonical user prompt as immutable gospel and physically locks later-stage agents to Read+Edit tools only, making full regeneration mechanically impossible. It is essentially a bureaucratic harness for LLMs: the creativity happens in the early steps, after which surgical patches, lint gates, and schema-enforced provenance try to prevent the system from quietly drifting off brief.

Key highlights

• Tiered depth modes: a light tier runs 5 steps in ~30–40 minutes, while full executes all 16 steps including contradiction graphs and parallel adversarial critics in ~1.5–2.5 hours.
• Persistent vault architecture: markdown notes are the source of truth, SQLite is a disposable cache; the vault supports full-text search, backlink graphs, and PDF ingestion.
• Subagent roster maps specific models (Haiku, Sonnet, Opus) to narrow roles like fetcher, loci analyst, dialectic critic, and patcher.
• Academic-first search strategy: it hits Semantic Scholar, arXiv, OpenAlex, and PubMed before touching general web search.
• Authenticated crawling via a local browser session for sources behind logins.

Caveats

• The claimed top spot on the DeepResearch-Bench leaderboard is an internal, forward-looking projection; third-party validation is explicitly noted as pending.
• It is tightly coupled to Anthropic’s Claude Code and model family (Opus, Sonnet, Haiku); running it requires both Claude Code access and budget for long, multi-agent sessions.
• The tool enforces structural hygiene (scaffold checks, provenance links, patch-surgery lint) but does not guarantee factual accuracy.

Verdict

Worth exploring if you already live in Claude Code and want a methodical, auditable research workflow that leaves a paper trail. Skip it if you need quick answers, cheap inference, or a tool-agnostic solution.