Skip to content
← Blog
5 min read

Teach Your Agent to Stop Hallucinating

One command installs two skills that teach any AI agent when to verify claims and when to use live data instead of guessing. Works with Claude Code, Cursor, Copilot, and Codex.

Every AI agent has the same problem: it doesn't know what it doesn't know. Ask it for a stock price and it'll give you one — from six months ago. Ask about earnings and it'll sound confident while citing numbers that were revised last quarter.

The fix isn't better prompting. It's teaching the agent when to look things up and when to check its work. That's what Agent Skills do.

bash
npx skills add veroq-ai/agent-skills

One command. Two skills. Works with Claude Code, Cursor, Copilot, and Codex.

What Agent Skills Are

Agent Skills are an open standard for teaching AI coding agents domain-specific knowledge. A SKILL.md file tells the agent not just what tools are available, but when to use them and why.

Supabase uses skills to teach agents about RLS policies. Stripe uses them for webhook patterns. We use them for something more fundamental: don't present unverified claims as fact.

Skill 1: veroq-verify

This skill triggers whenever your agent generates output containing prices, earnings, statistics, dates, or any verifiable statement. Instead of presenting claims and hoping they're right, the agent verifies first.

python
from veroq import shield

# Agent generates financial analysis
analysis = agent.run("Summarize NVIDIA's Q4 earnings")

# Verify before presenting to user
result = shield(analysis)

if result.claims_contradicted > 0:
    # Present corrections, not hallucinations
    for claim in result.claims:
        if claim.verdict == "contradicted":
            print(f"Correction: {claim.correction}")
else:
    # All claims verified — safe to present
    print(analysis)

The skill teaches the agent a decision tree: price claims go through the fast path (sub-300ms against live market data), SEC filings go through EDGAR, general claims go through the LLM verification pipeline, and high-stakes claims get cross-model consensus from six providers.

Every verification produces a permanent receipt with a public URL. Anyone can check veroq.ai/receipt/vr_abc123 and see the claim, verdict, evidence, and confidence score. That's the audit trail.

Skill 2: veroq-data

This skill triggers whenever the agent needs factual data — a price, a yield, a filing, a stat. Instead of guessing from training data, it calls VeroQ's live endpoints.

The default entry point is /ask — a natural language endpoint that auto-routes to the right data source:

python
from veroq import ask

# It figures out what you need
result = ask("What's happening with NVIDIA?")
# → price, signals, news, sentiment, earnings — auto-routed

result = ask("Is the yield curve inverted?")
# → full yield curve with inversion analysis

result = ask("What are Congress members buying?")
# → STOCK Act disclosure data

Behind /ask, there are 64 specialized endpoints covering:

  • •1,061 tickers with pre-computed signals refreshed every 5 minutes
  • •SEC EDGAR — filings, insider trades, XBRL financials
  • •Treasury yields with inversion detection
  • •Energy — WTI crude, natural gas, petroleum inventory
  • •Research — arXiv papers, FDA approvals, Congressional bills
  • •Social sentiment from X and Reddit
  • •Travel disruptions — TSA, FAA delays, border waits

All from free government data sources. The agent learns which endpoint to call for which question — and when to just ask /ask and let VeroQ figure it out.

Why Skills Matter More Than Tools

We already publish an MCP server with 78 tools. You can already pip install veroq and call any endpoint. So why skills?

Because tools tell agents what they CAN do. Skills tell agents what they SHOULD do.

An agent with 78 MCP tools available might still hallucinate a stock price because it doesn't know it should look it up. A skill that says “never guess a price — call fast/snapshot, it takes 300ms” changes the behavior at the decision level, not just the capability level.

Supabase found that “MCP alone is not enough” — agents knew how to implement security patterns but didn't know when. The same applies to verification. Agents can verify. They just don't unless you teach them to.

Install

bash
# Both skills — works with Claude Code, Cursor, Copilot, Codex
npx skills add veroq-ai/agent-skills

# Just verification
npx skills add veroq-ai/agent-skills@veroq-verify

# Just data
npx skills add veroq-ai/agent-skills@veroq-data

Your agent will start verifying claims and pulling live data in the next session. No configuration needed beyond a VeroQ API key.