deep-research skill — Usefulness & Uniqueness
Summary
- Useful for a specific audience: Developers who research inside coding workflows, need reproducible file-based artifacts, and want more auditability than consumer tools provide. Low value for fast ad-hoc lookups or users outside Cursor/Claude Code.
- Genuinely unique combination: No alternative ships all six patterns — IDE artifact delivery, per-claim atomic confidence with full URLs, parallel DAG subagents, plan pre-mortem, self-scoring patch loop, and SKILL.md portability — as one installable workflow spec.
- Individual patterns are academically validated (Flash-Searcher, PreFlect, DeepVerifier, LedgerMind, EviGraph) but no paper or tool combines them. The packaging and integration are the differentiator.
- Citation quality gap is real and growing: Haus Research audit (Sep 2, 2026) found 34.7% of Perplexity citation markers fail a numerical accuracy check. DRBench found Gemini hallucinates 13.3% of URLs. deep-research skill enforces full URL per claim.
- Market is consolidating: Microsoft retired consumer Copilot Deep Research (Aug 18, 2026) and is pivoting to MCP + agent patterns. If Copilot CLI
/researchadds structured evidence output, the gap narrows significantly.
Verdict
High for developers doing in-project research, competitive analysis, architecture decisions, or business research where output needs to live alongside code and be auditable. Low for fast ad-hoc lookups or users outside Cursor/Claude Code. Real adoption friction: file-based setup takes "a weekend at most." [stangeli.substack.com]
Real but integrative, not algorithmic. The differentiators are the per-claim evidence schema + pre-mortem critique + IDE-native file delivery as a combined single workflow. Weakest claim: SKILL.md workflows are ignored ~50% of the time when overlapping built-in model behaviors. [github.com]
Positioning risk: The market is moving toward composable MCP + agent patterns. Microsoft retired consumer Copilot Deep Research on Aug 18, 2026 [microsoft.com]. If Copilot CLI /research — currently hardcoded tools, no MCP access [github.com] — adds a structured evidence format, the gap closes. The pre-mortem phase and self-scoring loop are the hardest features to replicate without the full skill spec. Token overhead (~15×) and Gartner's projection that 40%+ of agentic AI initiatives will be cancelled by 2027 [joeyhipolito.dev] are real adoption risks.
Comparison table
| Dimension | deep-research skill | ChatGPT Deep Research | Perplexity Research | Copilot CLI /research | GPT-Researcher (OSS) |
|---|---|---|---|---|---|
| Runs inside IDE | Yes | No (web app) | No (web app) | Yes (CLI) | No (standalone) |
| Output as project files | Yes (timestamped dir) [src] | Partial (MD/Word/PDF) [src] | No (manual PDF only) [src] | Partial (~/ dir) [src] | Yes (Markdown) [src] |
| Per-claim confidence ratings | Yes (4-tier) | No | No (inline cites) | Partial (section) [src] | No |
| Atomic evidence store | Yes (evidence.md) | No | No | No | No |
| Full source URLs per claim | Yes (enforced) | No (10.1% non-resolving) [src] | No (34.7% fail) [src] | Partial (inline cites) | Partial (inline) |
| Parallel subagent DAG | Yes (explicit) | Unknown (internal) | Unknown (internal) | No | Yes (LangGraph) [src] |
| Plan pre-mortem phase | Yes | No | No | No | No |
| Self-scoring patch loop | Yes (0–12 rubric) | No | No | No | No |
| Browsable HTML dashboard | Yes (responsive) | No | No | No | No |
| SKILL.md portability | Cursor + Claude Code + 77 agents [src] | No | No | No (CLI only) | No |
| Open source | Yes (MIT) | No | No | No | Yes (MIT) |
| Typical run time | ~5 min | 5–30 min [src] | ~3 min [src] | Unspecified | Variable (high in deep mode) |
Key findings
Consumer AI research landscape (2026)
Five major consumer deep-research products (Perplexity, ChatGPT, Gemini, Claude, Grok) all keep output chat-bound with no structured per-claim evidence store. The Haus Research audit (Sep 2, 2026) of 1,826 Perplexity citation markers found 34.7% point to pages that don't open or contain none of the cited numbers. [hausresearch.com] DRBench benchmarks found Gemini Deep Research generates the most citations (113.1/query) but the highest hallucinated-URL rate at 13.3%; OpenAI achieves only 3.5% hallucinated URLs but a 10.1% non-resolving rate. [arxiv.org] Increased search depth degrades factual accuracy ~42% while link accessibility stays above 92%. [arxiv.org] Notably, OpenAI deprecated o3-deep-research and o4-mini-deep-research API models on July 23, 2026. [glasp.co]
Skills ecosystem (2026)
SKILL.md format open-standardized at agentskills.io (Dec 18, 2025), adopted by 77+ agent products including Cursor, Claude Code, VS Code, Gemini CLI, and GitHub Copilot. [agentskills.io] skills.sh (Vercel, Jan 20, 2026) is the discovery registry. [paperclipped.de] Deep-research install leaderboard (skills.sh API, Sep 3, 2026): firecrawl-deep-research 33,575 · parallel-deep-research 13,807 · 199-biotechnologies 9,675 · affaan-m/ecc 8,826 · samber/cc-skills 2,069. TheAdamLabs/deep-research-skill is not in the skills.sh index — no tracked install count. No official Cursor first-party registry; sharing is Git-based. [learncursor.dev]
Competing IDE-embedded tools
Copilot CLI /research is the closest first-party IDE competitor: saves cited Markdown to ~/.copilot/session-state/.../research/, exports as HTML/Gist, includes a "Confidence Assessment" section. [docs.github.com] Critical limitation: tool list is hardcoded (no user MCP servers) [github.com issue], defaults to claude-haiku-4.5, and produces no per-claim evidence file, no confidence-rated atomic store, and no HTML dashboard. MCP-based research servers (web-research-mcp) provide structured ResearchReport JSON with sub-question evidence and citation offsets, but leave full orchestration to the calling agent. [github.com] Microsoft retired consumer Copilot Deep Research on Aug 18, 2026. [microsoft.com]
Design novelty
All five design patterns have 2025–2026 academic validation: DAG parallel execution [Flash-Searcher] [AdaptOrch]; per-claim confidence scoring [HTC] [CEB]; pre-mortem planning (PreFlect achieves +8.5pp on GAIA over Reflexion) [arxiv.org]; self-eval loops (DeepVerifier +8–11% on GAIA) [aclanthology.org]; atomic evidence stores [LedgerMind] [LEDGER]. No surveyed paper or tool combines all five in one installable spec. [MisKnow-Agent] Confidence tiers are prompt heuristics — academic work uses calibrated ECE/AUC metrics. [arxiv.org]
Market gap
The clearest gap: structured, project-local, auditable research with per-claim full-URL evidence that lives alongside code. FS-Researcher confirms file-system workspaces let agents persist artifacts beyond the context window with negligible latency (<0.03% of wall-clock). [arxiv.org] Claude Code users explicitly demanded file-persistent research after silent transcript deletion (Jun 2026). [theregister.com] Gap is real but closable: MCP research servers and Copilot CLI updates could narrow it. Gartner projects 40%+ of agentic AI initiatives cancelled by 2027. [joeyhipolito.dev]
Evidence confidence
| Claim | Source | Freshness | Confidence |
|---|---|---|---|
| 34.7% of Perplexity citation markers fail a numerical accuracy check | hausresearch.com | Sep 2, 2026 | high |
| Gemini Deep Research hallucinates 13.3% of citation URLs; OpenAI 3.5% (DRBench) | arxiv.org/2604.03173 | Apr 2026 | high |
| Increased search depth degrades Fact Check accuracy ~42% while Link Works stays above 92% | arxiv.org/2605.06635 | May 2026 | high |
| MisKnow-Agent: one misleading doc raises false-conclusion adoption from 0% to 54.7% | arxiv.org/2607.20891 | Jul 2026 | high |
| PreFlect pre-mortem planning achieves +8.5pp on GAIA over Reflexion | arxiv.org/2602.07187 | Feb 2026 | high |
| DeepVerifier self-scoring loop delivers +8–11% accuracy on GAIA without training | aclanthology.org ACL 2026 | ACL 2026 | high |
| Flash-Searcher DAG parallel reduces steps 35% (11.2→7.4); 67.7% BrowseComp | arxiv.org/2509.25301 | Sep 2025 (ICLR 2026) | high |
| Microsoft retired consumer Copilot Deep Research on Aug 18, 2026 | support.microsoft.com | Aug 2026 | high |
| Copilot CLI /research cannot access user-configured MCP servers (hardcoded tools) | github.com/copilot-cli/issues/4076 | Jul 2026 | high |
| SKILL.md format adopted by 77+ agent products; skills.sh launched Jan 20, 2026 | agentskills.io | 2026 | high |
| Firecrawl-deep-research leads deep-research install leaderboard at 33,575 installs | skills.sh API | Sep 3, 2026 | high |
| Tow Center: AI search engines fail on 60%+ of queries; Perplexity best at 37% wrong | cjr.org (Tow Center) | Mar 6, 2025 | high |
| Multi-agent research uses ~15× more tokens than ordinary chat | anthropic.com | Jun 2025 | high |
| FS-Researcher: file-system workspace has negligible latency (<0.03% of wall-clock) | arxiv.org/2602.01566 | 2026 | high |
| No surveyed paper combines all 5 patterns (DAG, confidence, pre-mortem, self-eval, evidence store) in one integrated spec | arxiv.org synthesized | Sep 2026 | medium |
| samber/cc-skills deep-research at 2,069 installs on skills.sh (Sep 2026) | skills.sh API | Sep 3, 2026 | high |
| SKILL.md workflows ignored ~50% of the time when overlapping built-in behaviors | github.com/anthropics/claude-code/issues/30387 | Mar 2026 | high |
| Gartner projects 40%+ of agentic AI initiatives cancelled by 2027 | joeyhipolito.dev | Mar 2026 | medium |
| TheAdamLabs/deep-research-skill has no public install count or skills.sh listing | github.com/theadamlabs | Sep 2026 | medium |
| Confidence tiers are prompt heuristics not calibrated against ECE/AUC benchmark metrics | arxiv.org/2601.15778 (inference) | Jan 2026 | conflicted |
Sources (top 25 by importance)
- 1https://hausresearch.com/reports/perplexity-citation-audit/
- 2https://arxiv.org/html/2604.03173
- 3https://arxiv.org/html/2605.06635
- 4https://arxiv.org/abs/2607.20891 (MisKnow-Agent)
- 5https://arxiv.org/abs/2602.07187 (PreFlect)
- 6https://aclanthology.org/2026.findings-acl.1243.pdf (DeepVerifier)
- 7https://arxiv.org/abs/2509.25301 (Flash-Searcher)
- 8https://arxiv.org/abs/2601.15778 (HTC calibration)
- 9https://arxiv.org/abs/2606.07299 (DuMate-DeepResearch)
- 10https://docs.github.com/en/copilot/concepts/agents/copilot-cli/research
- 11https://github.com/github/copilot-cli/issues/4076
- 12https://support.microsoft.com/en-US/microsoft-copilot/deep-research-in-microsoft-copilot
- 13https://agentskills.io/specification
- 14https://www.skills.sh/api/search?q=deep-research&limit=20
- 15https://github.com/infinit3labs/web-research-mcp
- 16https://arxiv.org/html/2602.01566v2 (FS-Researcher)
- 17https://www.cjr.org/tow_center/…-ai-search-engines-theyre-all-bad… (Tow Center)
- 18https://www.anthropic.com/engineering/multi-agent-research-system
- 19https://arxiv.org/abs/2602.16873 (AdaptOrch)
- 20https://arxiv.org/abs/2607.28374 (LedgerMind)
- 21https://arxiv.org/abs/2608.18398 (LEDGER)
- 22https://github.com/anthropics/claude-code/issues/30387
- 23https://www.theregister.com/…/claude-code-users-complain-their-chat-records-are-being-mysteriously-wiped-out/
- 24https://www.digitalapplied.com/blog/open-source-deep-research-agents-2026-guide
- 25https://stangeli.substack.com/p/building-a-research-environment-that
Gaps & follow-up
What couldn't be found
- No independent end-to-end benchmark comparing deep-research skill vs. Copilot CLI
/researchon the same query. TheAdamLabs/deep-research-skillis not indexed on skills.sh — no tracked install count or adoption data.- Claude Code plugin community overlap with deep-research skill methodology not fully enumerated.
- Confidence tiers are prompt heuristics with no calibration against ECE/AUC metrics.
- No data on whether developers actually prefer file-based research artifacts over chat-mode output in practice (no 2026 survey).
Follow-up questions
- How does evidence quality compare to Copilot CLI
/researchon the same query, end-to-end? - Should the confidence tiers be backed by automated URL reachability + snippet verification to move from prompt heuristics to real calibration?
- Is there a path to skills.sh indexing for discoverability on par with firecrawl-deep-research (33,575 installs)?