A portable AI skill for patients

Critical AI
Health Literacy

An analytical lens for any topic where AI and health intersect. It asks the question most evaluations skip: who does this AI actually serve, and does it expand or constrain patient agency?

Download the skill See how it works

What it is

A way of seeing, not a fixed format

The framework comes from Hugo Campos and Liz Salmi's 2025 National Academy of Medicine commentary, "Critical AI Health Literacy as Liberation Technology," which applies Paulo Freire's theory of critical literacy to health AI.

You control what comes out, a quick take, a deep analysis, talking points, a post. The skill brings the lens. Most AI in healthcare is deployed by institutions to serve institutional goals. This skill helps you read past the marketing to the power underneath.

Read the paper →

What it does

Four dimensions, before any verdict

When the skill is active, Claude analyzes a tool, a study, or a news item along four dimensions instead of sorting it into a vague "good" or "bad" bucket.

  1. Who is the primary user?Patient-facing (used by patients, families, care partners) vs. clinician-facing (serves providers, like ambient scribes).
  2. Where is it hosted?Institutional, inside a health system or insurer, vs. public-facing, hosted for direct public use.
  3. Whose interests does it advance?The alignment question, and the one most people collapse by accident.
  4. Does it expand or constrain agency?A continuous scale from broadening patient capability to channeling people down predetermined paths.

The distinction at the center

Patient-directed ≠ patient-aligned ≠ patient-facing

Patient-directed is the goal: the agency-expanding end of the scale, AI that broadens a patient's capability to understand, decide, and advocate.

Patient-aligned is not the same thing. It means the interests behind the AI prioritize patients, though the AI may serve multiple stakeholders simultaneously and patient decision-making capacity varies. Alignment is not agency.

Patient-facing is weaker still. It only means the AI is intended for patients, families, or care partners, and says nothing about whose interests it serves or whether it expands agency.

Holding these apart is what lets the CAIHL lens read any situation with precision, whether it is a product, a paper, an article, or a post.

In practice

What that changes

Finer verdictsIt can say "patient-facing but not patient-aligned, public-facing, and agency-constraining," which a single axis can't express.
It critiques evaluationsHand it a study and it interrogates the method: single-turn prompts vs. real iterative use, clinician gold standards vs. patient-defined outcomes, and whether patients helped design it.
It reasons about orchestratorsAs personal AI assistants begin choosing tools on a patient's behalf, the same four questions apply to that hidden selection layer.
It keeps the Freirean coreReflection, strategic action, and agency, plus algorithmic resistance and AI as thinking partner rather than oracle.

How to use it

Point it at anything

Give it something to read, a news article, a study, a product page, a clinical note, a LinkedIn post, and ask for the CAIHL view. It shapes the output to whatever you need: a quick gut-check, a deep analysis, talking points, or a draft post.

  • Analyze this article from a CAIHL perspective.
  • Is this app patient-directed, patient-aligned, or just patient-facing?
  • Critique this study's evaluation method through the critical AI health literacy lens.
  • Help me think through an insurance appeal as an AI patient.

Does it work with any AI?

Yes. The framework is just text, so it runs in any assistant. Only the container changes.

Claude: packaged as a Skill, so it loads automatically. The one-click option.

ChatGPT: make a Custom GPT (or a Project) and paste the skill into its instructions. Reusable and shareable.

Gemini: create a Gem with the skill as its instructions.

Anything else: paste the skill into your chat as context, then ask your question. You lose the automatic triggering, but the lens works the same.

Install

Easiest in Claude, works anywhere

Claude is the smoothest path: it's packaged as a Claude Skill, so it installs in one upload and loads automatically. Any other assistant works too, see "Does it work with any AI?" above.

Claude apps

  1. Download the zip.
  2. Go to Settings → Skills → Upload zip.
  3. Ask something like: "Analyze this article from a CAIHL perspective."

Claude Code (or any folder-based agent)

git clone https://github.com/hugooc/caihl-skill
	mkdir -p ~/.claude/skills
	cp -r caihl-skill/critical-ai-health-literacy ~/.claude/skills/

Start a fresh session and run /skills to confirm it loaded.