From deep-research
This skill should be used when producing any research output, verifying claims from web sources, checking citation accuracy, assessing confidence in findings, preventing hallucination cascading across agent boundaries, or reviewing research documents for factual reliability. Covers the hallucination taxonomy (7 types), OWASP ASI08 cascading failures, circuit breaker patterns, citation verification rules, confidence scoring, ground-truth validation, and known limitations of automated verification.
npx claudepluginhub oborchers/fractional-cto --plugin deep-researchThis skill uses the workspace's default tool permissions.
Hallucination is the single most important engineering concern for research agents. A 5,000-word report with 100 claims at 5% hallucination probability per claim has a 99.4% chance of containing at least one hallucinated claim. Even the best-performing models hallucinate at measurable rates — 0.7% for simple summarization, rising to 5-13% on harder tasks (Vectara Hallucination Leaderboard, 2025...
Generates design tokens/docs from CSS/Tailwind/styled-components codebases, audits visual consistency across 10 dimensions, detects AI slop in UI.
Records polished WebM UI demo videos of web apps using Playwright with cursor overlay, natural pacing, and three-phase scripting. Activates for demo, walkthrough, screen recording, or tutorial requests.
Delivers idiomatic Kotlin patterns for null safety, immutability, sealed classes, coroutines, Flows, extensions, DSL builders, and Gradle DSL. Use when writing, reviewing, refactoring, or designing Kotlin code.
Hallucination is the single most important engineering concern for research agents. A 5,000-word report with 100 claims at 5% hallucination probability per claim has a 99.4% chance of containing at least one hallucinated claim. Even the best-performing models hallucinate at measurable rates — 0.7% for simple summarization, rising to 5-13% on harder tasks (Vectara Hallucination Leaderboard, 2025). In multi-agent systems, hallucinations compound across agent boundaries (OWASP ASI08).
Seven distinct hallucination types, ordered by detection difficulty:
| Type | What Happens | Detection | Prevention |
|---|---|---|---|
| Citation hallucination | Inventing papers, URLs, or authors that do not exist | Easy — verify URL/DOI exists | Never cite a source not actually retrieved and read |
| Temporal hallucination | Wrong dates or temporal ordering | Moderate — check against known timelines | Include dates from source text, not from memory |
| Factual fabrication | Entirely false statements presented as fact | Moderate — requires external lookup | Only state facts found in retrieved sources |
| Numerical hallucination | Fabricated statistics, percentages, counts | Hard — requires finding actual source | Copy numbers verbatim from source; never round or approximate without noting |
| Attribution hallucination | Real fact attributed to wrong source | Hard — requires cross-referencing | Track which source produced which claim |
| Negation hallucination | Reversing the polarity of a claim | Hard — requires careful reading | Quote or closely paraphrase source language |
| Conflation hallucination | Merging details from different sources into one false claim | Very hard — each component may be correct | Maintain per-source notes; do not blend findings until synthesis |
These rules are non-negotiable for all research output:
Prevent hallucination cascading across agent boundaries:
Treat each agent's output as untrusted input. When a research-worker agent returns findings, the synthesizing agent must not assume those findings are correct. Cross-reference key claims against other workers' findings or against the original sources.
Structured error responses. When a search or fetch fails, return an explicit error — never let an agent fill in missing data from "memory." A structured gap ("No information found on X") is infinitely better than a fabricated answer.
Validation gates between pipeline stages. Before synthesis begins, verify:
For citation formatting and management (inline format, Sources section structure), see the synthesis-and-reporting skill. This section covers verification of citation accuracy.
Core principle: Follow Perplexity's rule: "You are not supposed to say anything that you didn't retrieve."
Verification checklist for each citation:
Assign confidence levels to findings and communicate them in the output:
| Level | Criteria | Report Language |
|---|---|---|
| High | Claim appears in 2+ independent T1-T3 sources with consistent numbers (see source-evaluation skill for tier definitions) | State directly with citations |
| Moderate | Claim appears in 1 T1-T3 source or 2+ T4-T5 sources | "According to [Source]..." |
| Low | Single T4+ source, or sources partially conflict | "One source reports... though this could not be independently verified" |
| Unverified | No retrieved source supports the claim | Do not include, or explicitly flag as unverified |
Propagation rule: When combining findings from multiple agents, confidence of the combined finding equals the lowest confidence of its component claims.
Prefer deterministic validation over LLM-based validation:
| Check Type | Method | Example |
|---|---|---|
| URL existence | HTTP HEAD request | Verify cited URLs return 200 |
| Date verification | Parse and compare | Check if stated dates match source dates |
| Numerical consistency | String matching | Compare quoted numbers to source text |
| Cross-reference | Multi-source comparison | Same fact from independent sources |
LLM-based verification (asking a model "is this true?") is unreliable — models exhibit 17.8-57.3% bias-consistent behavior and high sycophancy rates (up to 58% initial compliance with wrong premises). Use code-based checks wherever possible.
The /research command enforces hallucination prevention through a three-stage pipeline:
This architecture addresses three hallucination failure modes:
Pipeline enforcement of confidence levels:
Things current automated systems cannot reliably verify:
For high-stakes research, human review of the final output remains essential.
For detailed hallucination research, OWASP ASI08 analysis, and verification architecture:
references/hallucination-research.md — Quantitative hallucination rates, cascading failure mechanics, AgentAsk error taxonomy, and multi-agent consensus patterns