Skip to content

arturseo-geo/grounded-research-skill

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

grounded-research-skill

Built by Artur Ferreira @ The GEO Lab · 𝕏 @TheGEO_Lab · LinkedIn · Reddit

Version Licence Claude Code PRs Welcome

Anti-hallucination research skill for Claude Code. Makes responses shorter and more conservative but dramatically more trustworthy.

Based on Anthropic's official hallucination reduction techniques.

Why This Exists

Claude is excellent at research — but default behaviour optimises for helpfulness, which sometimes means filling gaps with plausible-sounding content rather than admitting uncertainty. This skill flips the trade-off: accuracy over completeness, silence over speculation.

This is not a default mode. It significantly constrains creativity. Use it when factual accuracy is non-negotiable.

The Three Techniques

1. Admit Uncertainty

Claude gets explicit permission — and a strict obligation — to say "I don't know" rather than fabricate a plausible answer. Every gap in knowledge is surfaced, not papered over.

2. Extract Before Analysing

For any task involving documents or sources, Claude extracts word-for-word quotes first, then builds analysis anchored to those quotes. No analysis without evidence.

3. Cite Every Claim

Every factual statement must point to a source or get an explicit "I can't verify this" disclaimer. After generating a response, Claude runs a self-audit — any claim that fails verification is retracted publicly, not quietly removed.

Response Format

Every grounded research response follows this structure:

Section Purpose
Sources Consulted Every source used, with access method
Key Extractions Direct quotes organised by theme
Analysis Synthesis with inline citations to extractions
Confidence Assessment High / Medium / Low / Unknown per finding
Retractions Claims that failed self-audit, with explanation

When to Use

  • ✅ Fact-checked analysis
  • ✅ Document-based Q&A ("based on this document...")
  • ✅ Source-verified research
  • ✅ Citation-backed answers
  • ✅ Any task where accuracy > creativity

When NOT to Use

  • ❌ Creative writing or brainstorming
  • ❌ Code generation or debugging
  • ❌ Casual conversation
  • ❌ Exploratory research (use the research skill instead)
  • ❌ Opinion-based or speculative questions

Install

Manual install

git clone https://github.com/arturseo-geo/grounded-research-skill.git ~/.claude/skills/grounded-research

Via skills CLI

npx skills add arturseo-geo/grounded-research-skill

File Structure

grounded-research/
└── SKILL.md    # Complete skill — single file, always loaded

This is intentionally a single-file skill. No references directory, no complex structure. The entire technique fits in one document — adding more files would dilute the constraint.

How It Works Inside

The skill modifies Claude's behaviour through three complementary constraints:

┌─────────────────────────────────────────────────┐
│  1. UNCERTAINTY GATE                            │
│     "Do I actually know this?"                  │
│     If no → say "I don't know"                  │
│     If yes → proceed to extraction              │
├─────────────────────────────────────────────────┤
│  2. EXTRACTION LAYER                            │
│     Pull word-for-word quotes from sources      │
│     These become the ONLY foundation for        │
│     analysis — no quotes, no claims             │
├─────────────────────────────────────────────────┤
│  3. CITATION AUDIT                              │
│     Every claim → must have a source            │
│     No source found → retract publicly          │
│     Retraction is not failure — it's integrity  │
└─────────────────────────────────────────────────┘

Confidence Ratings

Level Meaning
High Multiple sources confirm, direct quotes support
Medium Single source, or inference from strong evidence
Low Limited evidence, reasonable inference but unverified
Unknown No evidence found — explicitly flagged

Contributing

This skill is deliberately minimal. Contributions should keep it that way.

Welcome:

  • Improvements to the constraint logic
  • Better examples of the technique in action
  • Edge case handling

Not welcome:

  • Adding reference files (keep it single-file)
  • Relaxing the constraints
  • Adding features that compromise accuracy for convenience

See CONTRIBUTING.md for guidelines.


Attributions & Licence

Built and maintained by Artur Ferreira · Part of the The GEO Lab toolkit.

Based on

All skill content is original writing. MIT licence.


Found this useful? ⭐ Star the repo and connect: 🌐 thegeolab.net · 𝕏 @TheGEO_Lab · LinkedIn · Reddit

Changelog

v1.0.0 (March 2026) — Initial release

Single-file skill implementing Anthropic's three hallucination reduction techniques: admit uncertainty, extract quotes before analysis, cite every claim with self-audit.

Related Repos

Licence

MIT — see LICENSE


Built and maintained by Artur Ferreira · Part of the The GEO Lab toolkit · MIT License

About

Anti-hallucination research skill for Claude Code — admits uncertainty, extracts direct quotes before analysis, cites every claim, retracts unverifiable statements. Based on Anthropic's official guardrail techniques. By TheGEOLab.net

Topics

Resources

License

Contributing

Security policy

Stars

26 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors