skills.rest
raian-pollock
raian-pollockCOMMUNITY
Explore all skills by this author →
GitHub

hallucination-reduction

Apply few-shot anchoring, temperature tuning, and Chain-of-Verification critics to reduce LLM hallucination.

Updated Jul 14, 2026
Search Tags:#prompt engineering#llm evaluation#hallucination reduction#citation verification#fact checking#chain-of-verification#few-shot prompting
One-click install
npx skills add https://github.com/raian-pollock/operator-skills --skill hallucination-reduction-raian-pollock
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: hallucination-reduction
Source: https://github.com/raian-pollock/operator-skills/tree/main/hallucination-reduction
Command: npx skills add https://github.com/raian-pollock/operator-skills --skill hallucination-reduction-raian-pollock

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? LLMs fabricate citations, statistics, named experts, and primary-source artifacts when generating factual content, and generic prompting advice does not measurably fix it. This Skill provides four validated pipeline techniques with measured lift so generated claims survive verification. ## Core Features & Use Cases - Four validated techniques: few-shot anchoring with diverse archetypes, temperature dial (0.3-0.4), Chain-of-Verification critic, and targeted repair, each with measured cost and lift data. - Content-type decision matrix: maps citations, primary sources, statistics, named experts, code, translations, and plans to the right technique combination, including when NOT to apply each technique. - Paste-ready templates: few-shot prompt suffixes and CoVe rubric JSON schemas per content type, plus documented failure modes like the schema-lock repair bug. - Use Case: When generating a bibliography for a research article, apply the citation few-shot suffix, run the citation CoVe rubric at temperature 0.1, and pipe flagged entries through CrossRef before publishing. ## Quick Start Ask the agent to apply the hallucination-reduction decision matrix and CoVe rubric before generating any content containing citations, statistics, or named sources.

Frequently Asked Questions about hallucination-reduction

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I reduce LLM hallucination when generating citations?▼

Combine few-shot anchoring with correctly formatted citation examples, lower temperature to 0.4, and run a Chain-of-Verification critic that checks author, year, journal, and DOI fields. CoVe alone cannot confirm a paper exists, so pipe flagged citations through CrossRef or Semantic Scholar before publishing.

What is Chain-of-Verification and how does it catch hallucinations?▼

Chain-of-Verification is a critic step where a model answers structured yes/no rubric questions about generated output at temperature 0.1, returning per-unit pass/fail JSON. It must use a different prompt path than the generator, since asking the same prompt to grade itself just re-rolls the dice.

Does lowering temperature always reduce hallucinations?▼

No. The validated win was at 0.4, not lower; drops below 0.3 collapse fluency in translation and creative tasks. Temperature is a structure-versus-diversity tradeoff, and few-shot anchoring contributes most of the measured fidelity gain.

When should I not use few-shot examples for factual generation?▼

Avoid few-shot for named expert quotes, since examples teach the model to write convincing fake quotes, and for code calling real APIs, where version-specific examples mislead. Also skip it for translations and creative copy, where it kills fluency and voice.

Why did a Chain-of-Verification repair step make output worse?▼

A documented failure mode: the targeted-repair merge accepted the model's value for a schema-pinned field, breaking a required structural constraint. The fix is to merge content fields from the repair but lock schema-pinned fields to their original values.

Related Skills

Explore More Software Engineering Skills→
udecodeudecode
434

task

Execute tasks end-to-end from context gathering to verification.

Official
Advanced
MoxyWolfLLCMoxyWolfLLC

citation-verifier

Verify academic citations via DOI, arXiv, and semantic relevance checks.

Official
Advanced
AncienttwoAncienttwo
407

claude-review

Review code changes with a different model before merging.

Community
Intermediate
CrestAppsCrestApps
13

orchardcore-ai-chat-interactions

Configure ad-hoc AI chat interactions for Orchard Core with CrestApps modules.

Official
Advanced