跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
搜索与研究

desk-research

Cited secondary research via two parallel sub-agents (competitive scan + scholar scan), both forced to surface disconfirming evidence. Use when the user wants desk research or a literature review, a competitive analysis of how others solved a problem, or asks 'is there evidence for' a hypothesis.

superhero-tech · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add superhero-tech/superhero-skills --skill desk-research

技能说明


name: desk-research description: "Cited secondary research via two parallel sub-agents (competitive scan + scholar scan), both forced to surface disconfirming evidence. Use when the user wants desk research or a literature review, a competitive analysis of how others solved a problem, or asks 'is there evidence for' a hypothesis." metadata: ready_to_publish: "true"

Desk Research

Frame the question with the user, design queries together, dispatch two parallel sub-agents (competitive + scholar), synthesize, and produce a report with mandatory disconfirming evidence.

Respond and produce all artifacts (including the report) in the user's conversation language.

This skill runs on live web search. Confirm web search tools are available before proceeding; if they are not, tell the user and stop - the report must come from live sources, not memory.

Reference load map

Load each file only at the stage indicated:

StageLoad
0 Setupassets/state-template.json
3 Dispatchassets/shared-subagent-base.md + assets/competitive-subagent-prompt.md / assets/scholar-subagent-prompt.md
4 Synthesisreferences/quality-rubric.md (re-check confidence, synthesis checks)
5 Reportreferences/report-template.md

Hard rules (always active)

  1. Confront the hypothesis, don't confirm it. At least 1/3 of the report must address disconfirming evidence. This is the heart of the skill.
  2. Every finding needs a URL. No URL = not a finding.
  3. Minimum 3 independent sources for High confidence. Three articles citing the same report = 1 source.
  4. Vendor sources flagged, not hidden. Confidence capped at Low, tagged [VENDOR] in bibliography.
  5. Never modify ost-state.json.

Full scoring criteria live in references/quality-rubric.md.

Lazy user mode

If the user signals disengagement ("skip", "", "." or equivalent): auto-decide, state your pick and why in one sentence, mark auto_decided: true in state, continue. Exception: if the disconfirmation prompt (1.4) is skipped, mark disconfirmation_skipped: true and strengthen the sub-agents' falsification mandate.

Workflow

Setup → Brief gate → Query design gate → Sub-agent dispatch (||) → Synthesis gate → Report

If invoked on existing state, jump to the stage in current_stage (see Resume routing).

Stage 0 - Setup

  1. Detect state. Look for desk-research-state-*.json in the project folder or research/. If found, offer: resume / jump to stage / start new (archive old).
  2. Ingest context. Scan the project folder: brief.md and *-brief.md; first 80 lines of every .md in research/ and analyses/; ost-state.json read-only (outcome statement + top 2-3 opportunities by PRES); note *.csv/*.xlsx filenames without loading them. Extract industry_or_domain from the brief; if the folder is empty, ask directly.
  3. Confirm scope (single-choice): single problem (default) / broader literature scan (scale=deep) / competitive only / scholar only. Initialize state from assets/state-template.json.

Stage 1 - Research brief gate

  1. From context, formulate: problem statement (one falsifiable sentence), working hypothesis (what the user already assumes, often unstated), decision at stake (what gets approved/rejected). Ask directly for anything you can't extract - don't invent.
  2. Propose 3-5 falsifiable research questions ("Does X correlate with Y?", "Under what conditions..."). Show a numbered list; loop with the user until approved.
  3. Set scale from question count: 1-3 → quick (~5-7 sources per sub-agent, 1-2 page report); 4-5 → deep (~15+ sources, full report). Tell the user explicitly, offer override.
  4. Disconfirmation prompt (soft gate). Ask what evidence would prove the hypothesis WRONG. Accept any answer; if skipped, mark disconfirmation_skipped: true.
  5. Show the full brief (problem, hypothesis, decision, questions, scale, falsification criterion). Gate: OK / edit / stop and discuss.

Stage 2 - Query design gate

  1. Competitive queries. Per research question: 5-10 companies/cases (leaders + niche + adjacent industry), 3-5 English search keywords, recency window (default 5 years), source heuristics (prioritize trade press, analyst reports, case studies; flag vendor whitepapers).
  2. Scholar queries. Map the user's business language to academic language. Show the mapping table (USER LANGUAGE → ACADEMIC LANGUAGE, 1-2 query strings each) - this is where users often add a key term.
  3. Confirm all queries with the user (edit/add/remove), save the final set to state.

Stage 3 - Sub-agent dispatch (parallel when available)

Dispatch both scans with the host's sub-agent or delegation capability. Start them in parallel when the host supports parallel delegation. In Claude Code, a general-purpose Agent is suitable; in Codex, use the available sub-agent/collaboration mechanism. Do not depend on one provider's tool name. If the host has no delegation capability, run the scans sequentially in the main context and keep their findings in separate files.

  • A - Competitive scan: prompt = assets/shared-subagent-base.md + assets/competitive-subagent-prompt.md.
  • B - Scholar scan: prompt = assets/shared-subagent-base.md + assets/scholar-subagent-prompt.md.

Substitute every {{placeholder}} from state (placeholder names match state fields). Computed: {{recency_start_year}}, {{current_year}}. Special: {{disconfirmation_criterion_OR_skipped_message}} - the criterion verbatim, or if skipped: "User skipped the falsification criterion. Your mandate for the disconfirming section is stronger: minimum 4 disconfirming findings at confidence Medium+."

Each sub-agent saves findings to {project_path}/research/.desk-tmp/<stream>-findings-<timestamp>.md and returns a 200-word summary.

Failure handling: competitive empty → queries too narrow, back to 2.1. Scholar empty → may be genuinely under-researched

阅读完整来源 ↗