super-research

Question in → tier-classified deep research → on-brand HTML report → hosted URL.

A Claude Code skill that chains two other skills (included) into one end-to-end research-publishing pipeline:

your question
   │
   ▼
super-research-    parallel research agents · source tier-classification (T1–T4)
   engine          cross-referencing · verification gates · confidence labels
   │               → markdown report on disk
   ▼
share-as-page      brand-token resolution (cached per domain) · self-contained HTML
   │               screenshot visual eval · zero external requests
   ▼
hosted URL         one-command Vercel deploy (surge/netlify fallbacks)

The three skills

Skill	Role
super-research	Thin orchestrator: scope → research → page → publish
super-research-engine	Multi-agent research with source tiers, contradiction resolution, and hard quality gates (every finding cited, failed angles disclosed, no fabricated URLs). Invoked by super-research; named distinctly so it never collides with Claude Code's built-in /deep-research workflow
share-as-page	Report → polished self-contained HTML wearing a real brand's tokens, visually verified via headless-Chrome screenshots before deploy

Each also works standalone.

Install

git clone https://github.com/robertnowell/super-research
cp -R super-research/skills/* ~/.claude/skills/

Optional but recommended — the deterministic trigger hook (skill discovery by description-matching is probabilistic; a hook isn't). Register skills/super-research/hooks/route-to-super-research.py as a UserPromptSubmit hook in ~/.claude/settings.json. It steers research prompts to super-research and away from Claude Code's token-heavy built-in /deep-research workflow.

Heads up: newly installed skills may need a fresh Claude Code session before they're callable — if /super-research isn't found, start a new session.

Use

In Claude Code:

/super-research <your question>

or just ask naturally: "deep research X and make me a page for ".

The skill will ask (at most once) about research depth, whose brand the page should wear, and whether the content is too sensitive for an auto-deploy.

Requirements

• Claude Code
• Chrome/Chromium (screenshot visual-eval gate)
• vercel CLI, logged in (deploy step; surge/netlify work as fallbacks)

Notes

• Pages are single self-contained HTML files: inline CSS, base64 logos, no external requests. They render double-clicked offline and print cleanly to PDF (⌘P).
• Brand tokens are cached per domain in share-as-page/brands/*.json after first resolution, so the second page for a brand is instant.
• The research report (markdown) is always preserved on disk as the source of truth; the page renders it and never silently adds facts.
• The research engine is deliberately named super-research-engine, not deep-research: the bare name deep-research collides with Claude Code's built-in /deep-research workflow (a separate, token-heavy primitive) that would shadow this skill and fire instead. Don't rename it back. As defense-in-depth, super-research and the engine both set disallowed-tools: Workflow.

Token footprint

super-research runs on Claude Code's Agent tool, not the Workflow tool — deliberately. A quick-scope run fans out 2 research agents at ~25k tokens each (~51k total, measured); a full run lands around ~100k. Claude Code's built-in /deep-research workflow, by contrast, routinely runs >3M tokens per run (measured, observed repeatedly). Same multi-agent research, 1–2 orders of magnitude less spend — which is also why the engine sets disallowed-tools: Workflow.

License

MIT