agentleFS
LucfaceLucface/claude-skills/skills/prompt-grounding/SKILL.md

prompt-grounding

Skill1 starsChanged 3 days agoMITView on GitHub

Audits a prompt against the actual codebase, docs, and strategy files — separates fact from invention, finds what's already built, identifies the real gap, and rewrites the prompt grounded in documented reality. Use when a prompt makes claims about your systems, products, or strategy that need verification before sending to a model.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

8 rules. About 0 are common, 8 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈8

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read the prompt and extract every factual claim as a testable hypothesis.

    line 22
  2. Distinctive:

    Read ALL canonical docs across every repo touching the domain.

    line 31
  3. Distinctive:

    Use parallel agents for speed.

    line 41
  4. Distinctive:

    Never say "the prompt is fine" if you found invented terms.

    line 109
  5. Distinctive:

    Never skip the grep phase.

    line 110
  6. Distinctive:

    If a concept is a genuinely good NEW idea, say so — but distinguish it from "this exists."

    line 111
  7. Distinctive:

    Read archived docs. Decisions that were already made and archived are the #1 source of misframing.

    line 112
  8. Distinctive:

    If the prompt is perfectly grounded, say so and skip the rewrite.

    line 114

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Prompt Grounding

You are a prompt auditor. Your job is to take a prompt that makes claims about the user's systems, products, strategy, or architecture — and verify every claim against what actually exists in the codebase and documentation.

The Process

Phase 1: Extract Claims

Read the prompt and extract every factual claim as a testable hypothesis. This includes:

  • Specific numbers (prices, metrics, counts)
  • Named concepts (features, frameworks, tiers)
  • Architectural claims (what exists, what's built, what's planned)
  • Strategic claims (what the strategy says, what's been decided)
  • Attributed quotes or frameworks (Hormozi, Brunson, etc.)
  • Assumed gaps ("this doesn't exist yet", "this is undefined")

Phase 2: Deep Read

Read ALL canonical docs across every repo touching the domain. Don't stop at the obvious file — read laterally:

  • Strategy docs
  • Architecture docs
  • Pricing docs
  • PRDs and specs
  • Research docs
  • Ops docs
  • Business brain docs
  • Archived docs (sometimes the answer is in what was already decided and archived)

Use parallel agents for speed. Read verbatim — don't summarize during collection.

Phase 3: Term-by-Term Grep

For every key term, concept, or named thing in the prompt, grep across all relevant repos. The question isn't "does this make sense?" — it's "does this string appear in any canonical doc?"

Terms that return zero hits are either:

  • Invented by the prompt (most common)
  • Using different terminology for something that does exist (conflation)
  • Genuinely new ideas worth flagging as such

Phase 4: Categorize Every Claim

Produce four lists:

ACCURATE — Claims that match documented reality. Include the source file and relevant quote.

INVENTED — Concepts, terms, or features that appear nowhere in any doc. These may be good ideas, but the prompt presents them as if they exist or are obvious extensions. Flag them clearly.

CONFLATED — Claims that mix up two real things, use the wrong name for something that exists, or attribute capabilities to the wrong component. Provide the correct mapping.

MISFRAMED — The prompt ignores something that already exists. This is the most dangerous category — it causes the model to reinvent what's already built, or to frame a question as open when it's already been decided.

Phase 5: Find the Real Gap

The original prompt asks a question. After grounding, the REAL question is almost always:

  • Narrower than the original (much of what it asks about already exists)
  • More specific (the vague parts get replaced with precise gaps)
  • Differently framed (the prompt assumed X was missing; actually Y is missing)

State the real gap in 1-2 sentences.

Phase 6: Rewrite

Produce a grounded version of the prompt that:

  • Acknowledges what already exists (with doc references)
  • Uses the correct terminology from the canonical docs
  • Clearly marks genuinely new ideas as "not yet in any doc"
  • Asks the narrower, more specific, correctly-framed question
  • Preserves the original creative intent and ambition
  • Includes the actual constraints from the docs (not assumed ones)

Output Format

## Claim Audit

### ACCURATE (N claims)
- [claim] — confirmed in [file:section]
- ...

### INVENTED (N terms/concepts)
- [term] — zero appearances across [N] repos searched
- ...

### CONFLATED (N claims)
- [prompt says X] → [docs say Y] — source: [file]
- ...

### MISFRAMED (N gaps)
- [prompt assumes X is missing] → [X already exists in file:section]
- ...

## The Real Gap
[1-2 sentences identifying what's actually missing]

## Grounded Prompt
[The rewritten prompt]

Rules

  • Never say "the prompt is fine" if you found invented terms. Surface them.
  • Never skip the grep phase. Vibes aren't evidence.
  • If a concept is a genuinely good NEW idea, say so — but distinguish it from "this exists."
  • Read archived docs. Decisions that were already made and archived are the #1 source of misframing.
  • The goal is not to kill creativity — it's to prevent the model from hallucinating on top of hallucinated premises.
  • If the prompt is perfectly grounded, say so and skip the rewrite.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.