auto-research-agent

name	auto-research-agent
description	Deep research runs with explicit source policy, comparison tables, confidence levels, Q&A appendix, and gap analysis—outputs a dated research pack under workspace/drafts/research/. Karpathy-style: one primary artifact per run, clear done criteria.

auto-research-agent

Run autonomous research with traceability: every non-obvious claim should tie to a source note (URL, doc title + date, or “model inference—low confidence”). Output is a single primary markdown report per run (plus optional CSV), same “one dated artifact” discipline as qf-course-researcher uses for Notion.

Prerequisites

• Workspace: read USER.md / SOUL.md; write under:

  workspace/drafts/research/<YYYY-MM-DD>-<slug>/
  

• Tools: web search / browse when available. If unavailable, entire report must be labeled INFERENCE-ONLY in H1 banner and no factual “trending” or stats claims.
• Scope: timebox (e.g. 45 min equivalent depth), geography, language from human.

Credentials & API (qf-style)

• Browse / web search only: No files required if the gateway exposes search tools (same class as qf-course-researcher “Web access”).
• Optional structured search: If you add SerpAPI, Brave Search API, etc., put keys in ~/.openclaw/.env and document the variable names in workspace/INTEGRATIONS.md (add a row to the table there when you introduce one).
• QuizFactor cross-check: If the brief needs live course data, reuse qf-course-researcher patterns: QF_BASE_URL from ~/.config/quizfactor/base_url and GET /api/ai/courses with Bearer if your API requires it — see workspace/INTEGRATIONS.md and qf-course-researcher Prerequisites.

High-level Workflow

1. Scope contract (00-scope.md)

   • Questions answered, out of scope, success criteria, stakeholder (who uses this).

2. Source plan

   • List source types you will use: official docs, filings, academic, news (with dates), forums (low weight).
   • Set minimum source count (e.g. ≥8 for deep dive, ≥3 for quick brief).

3. Gather — structured notes (notes-raw.md)

   • Bullet per source: Source: URL or citation, Date accessed:, Claim supported:, Confidence: high/med/low.

4. Synthesize — executive summary (report.md)

   • TL;DR (5 bullets max).
   • Key findings with inline source markers [S1], [S2] mapping to a Sources section at bottom (numbered list).

5. Comparison tables

   • When comparing options (tools, vendors, courses), add comparison.csv or markdown table:
     • Columns: Name, Pros, Cons, Cost signal, Risk, Source.

6. Q&A appendix

   • qa.md: anticipated stakeholder questions → short answer + pointer to section in report.md.

7. Gaps and contradictions

   • Section Gaps: unknowns, conflicting sources, what would change conclusion if known.

8. Recommendations

   • Next steps numbered; each actionable (who/what/when style, no vague “monitor”).

9. Handoff

   • README-handoff.md: which downstream skill consumes this (product-manager-agent, social-trend-monitor, qf-course-researcher-style workflows).

10. Scheduling

    • Daily pulse research: short slug daily-<topic>.
    • Deep dive: slug deep-<topic>; do not overwrite prior dated folders.

Outputs (required)

• workspace/drafts/research/<YYYY-MM-DD>-<slug>/report.md (primary)
• 00-scope.md, notes-raw.md, README-handoff.md

Optional: comparison.csv, qa.md

Agent Checklist

• Scope file created first; stuck to scope or documented deviation.
• INFERENCE-ONLY banner if no web/tools access.
• No fabricated statistics, dates, or rankings.
• Source list numbered; report references match.
• Gaps section non-empty when data is incomplete.
• Folder path communicated to user.
• Recommendations tie to business/product/social next actions.