The Grounded Research Skill.

A free Claude Skill that grounds every fact in your research to a real source. Built for moments when being wrong has a cost — discovery calls, vendor comparisons, regulation lookups.

Free · sent by email

We email the ZIP plus a 4-step install. Unsubscribe whenever.

grounded-research.skill

# Every fact traces to a source.name: "grounded_research"version: 1.1invoke: "/grounded-research [topic]" # Discipline (non-negotiable)rules:- WebSearch for URLs only- WebFetch is the citation path- No training-data fallback- Verbatim excerpts logged- Stale sources flagged- Honest gaps reported audit_trail: "core/sources.json"on_save: append_with_excerpt()

Stop trusting the confident-sounding answer.

Default Claude is helpful for casual questions. It's also confidently wrong sometimes — paraphrased search snippets that turn out to be misquotes, statistics with no source, "according to a 2023 study" that doesn't exist. For a casual answer that's annoying. For a discovery call where you're claiming a competitor charges $X, it's career damage.

The Grounded Research Skill fixes that. Install it once, and every fact in every research brief traces to a real URL Claude actually fetched, with the exact quote logged in a permanent audit trail.

Every claim survives a fact-check, or it doesn't get into the brief.

Who it's for.

Owners who need research they can defend. Use it when prepping for a discovery call and you need facts on the prospect's industry, comparing vendors and you need real pricing, or looking up regulation, policy, or contract-clause specifics.

Don't use it for casual questions. Plain Claude is the right answer there.

What you get.

The Skill install file (grounded-research-v1.1.zip).
A research playbook that fetches every source, logs verbatim excerpts, and refuses to fabricate when a fetch fails.
A permanent audit trail at core/sources.json — every URL, every quote, every retrieval method, dated.
An email if we update the Skill or Anthropic changes how Skills work.

How to install.

Quick. Five minutes once the email arrives.

Install sheet5 min

Download the install file from the email we send.

Arrives within a minute or two of submitting the form above.

In Claude Cowork, open Customize → Skills → + → Create skill → Upload a skill.

Pick the ZIP from the email. Custom Skills need code execution enabled on your plan.

Run it with `/grounded-research [topic]` in any Cowork task.

Claude switches to the grounded playbook for that conversation.

Read the brief at `core/research-<topic>.md`.

Every fact ends with [U042] mapping to the ledger. Audit any claim.

Module 5 has screenshots

What it's not.

Not an accuracy upgrade for every Claude conversation. Skills are content, not interceptors. To apply the discipline to research done by other skills, paste the house-rule text (in the skill's README) into your project's Custom Instructions.
Not magic. If the source you need is paywalled or anti-bot-blocked, the skill won't pretend — it'll honestly report the gap. That's the feature.
Not auto-invoked. You have to ask for it: /grounded-research [topic] or "use the grounded-research skill to look up X." This is deliberate.
Not a substitute for your judgment. Citations make verifying fast — but verification is still your job. LLMs can quote imprecisely, attribute to the wrong line, or fetch a source that itself contains an error. The skill's job is to make every claim auditable. Auditing is yours.

Get the Skill.