Ground first, ask second

rule Thu Jul 02 2026 00:00:00 GMT+0000 (Coordinated Universal Time) promptingagentgroundinghallucination

Ground first, ask second

The rule

When a task needs specific domain knowledge, up-to-date info, or a particular framework/style, split into TWO prompts in the same session:

1. Ground prompt: "Research X. Extract the important tactics/patterns/spec. Be up-to-date."
2. Action prompt: "Now using what you just researched, do Y."

Same chat session. Grounding lands in context; action prompt draws from it.

Example

Task: "Help me launch this product."

❌ Single-prompt:

Help me launch this SaaS product. What should my pricing strategy be?

→ Generic pricing advice.

✅ Ground first:

Prompt 1: Research Alex Hormozi's product-launch playbook. Extract:
  value equation, Grand Slam Offer components, pricing philosophy,
  market selection rules, risk reversal tactics. Be thorough.

Prompt 2 (same chat): Now using those tactics, help me price + position
  my SaaS. Product details: [full context].

→ Grounded, specific advice referencing the real playbook.

When to use

• Task references a specific methodology / framework / expert
• Task involves fresh info (recent library version, new API, current market)
• Task in a domain you're not sure Claude is current on
• Iterating on a design that references specific artistic movements / brands

When to skip

• Simple mechanical task
• Task where Claude's baseline knowledge is definitively sufficient
• Task with all context already in knowledge/ (just reference the file)

Grounding methods

Ground prompt can:

• Research on the web (Claude's web search): "research X"
• Draw from uploaded files: "read this doc + summarize the key tactics"
• Reference internal knowledge: "load knowledge/decisions/stack/pipeline-stack-2026-07-01.md before answering"
• Combine: "read this doc AND search web AND cite both"

Combines with ICC

Ground prompt IS an ICC prompt: instructions ("research and extract"), context ("this is grounding for a later prompt about Y"), constraints ("extract as bullet list, cite sources").

Anti-patterns

• ❌ Ground and act in one prompt — model tries to shortcut and does both poorly
• ❌ Ground in one session, act in another — context lost, defeats the pattern
• ❌ Ground with a shallow one-liner — get shallow grounding, useless downstream
• ❌ Skip grounding on fresh-info tasks — get 2024-training-cutoff answers

Cross-refs

• icc-prompt-formula — the shape of both the ground and action prompts
• always-search-twice-before-deciding — "search twice" applies to grounding too
• context-interview — alternative when the gap is user-context not domain-context
• Source: Nate Herk YouTube 2026-07 "Ground first, ask second" pattern