Ground Truth: Critical Thinking Partner

An agent skill for calibrated honesty. Ground Truth activates a critical thinking partner with hard anti-sycophancy rules: the advisor who will actually tell you the truth.

Why this exists

AI assistants are trained to be helpful, which in practice often means agreeable. You push back on a claim, they fold. You float a half-baked idea, they validate. You ask for a review, they lead with what's working.

This skill fixes that. Not by making an assistant disagreeable -- that's just a different kind of useless -- but by enforcing calibrated confidence: say what's actually true, hedge only when genuinely uncertain, and name the problem with an idea instead of working around it.

Before and after

You ask: "I am thinking of launching this product next week. The landing page looks polished and a few friends said they would use it. Does that hold up?"

Default assistant: "That sounds promising. A polished landing page and positive feedback are strong early signals."

With Ground Truth: "The weakest assumption is that friendly feedback predicts demand. It does not. Before launching, get evidence from people who are not socially invested in encouraging you: sign-ups, deposits, or at least problem interviews with a clear commitment signal. What would count as disconfirming evidence for you?"

Install for Claude

Download ground-truth.skill from the latest release, then add it through your Skills settings.

Install for Codex

Ask Codex:

Install my Ground Truth skill from https://github.com/glichtenthal/ground-truth

Or install manually:

python3 ~/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py \
  --repo glichtenthal/ground-truth \
  --path . \
  --name ground-truth

Restart Codex after installation.

Manual install

The repository is the skill. Copy the complete folder into your agent's skills directory. For example, Codex discovers:

~/.codex/skills/ground-truth/SKILL.md

What it does

Use Ground Truth when you want honest assessment instead of validation on plans, decisions, ideas, or work in progress. Default mode is calibrated honesty, not reflexive pushback.

The full behavior rules, sycophancy patterns to block, opener blacklist, and tone guidelines are in SKILL.md.

Version history

See CHANGELOG.md.

More agent skills

• The Briefing Room - turns messy context into a structured brief a human can think with.
• The Quorum - a five-member expert council that pressure-tests consequential decisions from multiple angles.