ground-response

/prompt-engineering:ground-response

Produce a grounded, citation-backed response — every claim traced to a source quote, unsupported claims retracted, unknowns stated explicitly.

When to Use This Skill

Use this skill when...	Use something else when...
You need factual accuracy grounded in source documents	Text needs style/tone adjustment → /prose:distill
Analyzing long documents (>20k tokens) where hallucination risk is high	You need code review → /code-quality:code-review
User asks to "cite sources" or "verify claims"	You need to synthesize scattered notes → /prose:synthesize
Answering questions about specs, policies, or technical docs	General coding task with no source verification needed
User wants auditable, traceable answers	Creative writing or brainstorming (accuracy not the goal)

Core Principles

These three techniques are from Anthropic's official documentation on reducing hallucination. Apply all three in every response.

1. Permit Uncertainty

You have explicit permission to say "I don't know" or "The source does not address this." Do not invent, speculate, or fill gaps with plausible-sounding information. When the source is silent on a topic, say so. When the source is ambiguous, state the ambiguity.

2. Extract Direct Quotes First

Before analyzing or answering, extract word-for-word quotes from the source material that are relevant to the task. This grounds your reasoning in actual text, not recalled impressions. For documents >20k tokens, this step is critical — extract before you analyze.

3. Verify Claims Against Quotes

After formulating your response, audit every claim. Each claim must have a supporting quote. If a claim lacks a supporting quote, retract it. Mark retracted claims with [RETRACTED — no supporting quote found]. This makes your response auditable.

Parameters

Parse $ARGUMENTS:

• Question/task: The primary argument — what to analyze or answer
• --source <path>: Optional file path or glob pattern for source material
• If no --source is provided, look for source material in the conversation context
• If no source material is available anywhere, ask the user to provide it

Execution

Execute this grounded analysis workflow:

Step 1: Identify and Read Source Material

1. If --source is provided, read the file(s):
   • Single file: use Read
   • Glob pattern: use Glob to find files, then Read each
   • Directory: use Glob with **/* to discover relevant files
2. If no --source, check conversation context for documents, code, or prior content
3. If no source material is available, stop and ask: "What document or source should I ground my response in?"

Note the total size of source material. For large sources (>20k tokens), be especially rigorous about Step 2.

Step 2: Extract Direct Quotes

Search the source material and extract word-for-word quotes relevant to the question/task.

For each quote, record:

• Quote number: Sequential identifier (Q1, Q2, etc.)
• Exact text: Word-for-word from the source — no paraphrasing
• Location: File path and line number, or document section

Extract comprehensively. It is better to extract too many quotes than too few — you can discard unused quotes later, but you cannot cite quotes you didn't extract.

If no relevant quotes are found, state: "No relevant quotes found in the provided source material. I cannot provide a grounded response to this question."

Step 3: Formulate Claims

Answer the question or perform the task. Structure your response as a series of claims, each referencing one or more quotes:

• Tag each claim with its supporting quote(s): [Q1], [Q2, Q3]
• Keep claims atomic — one assertion per claim
• Distinguish between what the source explicitly states and what you infer

Step 4: Verify — Audit Every Claim

Walk through each claim in your response:

1. Does it have a tagged quote reference?
2. Does the referenced quote actually support this specific claim?
3. Is the claim a faithful representation of the quote, or does it overstate/distort?

For any claim that fails verification:

• If the claim overstates the quote → weaken it to match what the quote supports
• If no quote supports it → mark as [RETRACTED — no supporting quote found]
• If the quote is ambiguous → note the ambiguity explicitly

Step 5: State Unknowns

After verification, explicitly list:

• Questions the source does not address — gaps in the material relevant to the task
• Ambiguities — where the source is unclear or contradictory
• Limitations — boundaries of what can be concluded from this source alone

Do not skip this step. Stating unknowns is as important as stating knowns.

Step 6: Present Grounded Response

Deliver the final response using this structure:

## Grounded Analysis

<Your response with inline [QN] citations for each claim>

## Supporting Quotes

| # | Quote | Source |
|---|-------|--------|
| Q1 | "exact text from source" | file.md:42 |
| Q2 | "exact text from source" | file.md:78 |

## What the Source Does Not Address

- <Gap 1>
- <Gap 2>

## Retracted Claims (if any)

- <Claim that was removed and why>

Handling Edge Cases

Situation	Approach
Source contradicts itself	Quote both passages, note the contradiction, do not resolve it
Question is partially answerable	Answer the answerable part with citations, list unanswerable parts in unknowns
Source is code, not prose	Extract relevant code blocks as quotes, cite file:line
Multiple sources disagree	Present each source's position with its quotes, note the disagreement
Source is very short (<1k tokens)	Still follow the full workflow — brevity doesn't eliminate hallucination risk

Agentic Optimizations

Context	Approach
Short document (<5k tokens)	Read fully, extract quotes inline, present compact response
Long document (5k-50k tokens)	Use Grep to find relevant sections, then Read targeted ranges
Very long document (>50k tokens)	Use Grep with multiple patterns, read only matching regions
Codebase analysis	Use Glob + Grep to locate relevant files, extract code quotes with file:line
Multiple files	Process each file, merge quotes, deduplicate overlapping citations
No source provided	Ask user before proceeding — never ground in imagined sources