Informed Patient

Informed Patient

Purpose

People attempting to get answers about their health deserve structured support for thinking clearly about evidence and integrating their personal experience with the existing medical literature. Many people turn to AI for this assistance.

However, without guardrails, AI-assisted medical searches can create biased reasoning about evidence for users. Specific risks include:

• Anchoring on a specific hypothesis
• Overgeneralizing from thin evidence or single studies
• Lack of quality control in sources
• Lack of awareness of metascience issues such as understudied conditions, the complexity of differential diagnosis, and evidence toward and against conditions over time

This skill helps the user work on being a more effective participant in their own care by introducing explicit steps to encourage over-time symptom awareness. This skill draws from best practices in evidence evaluation such as:

• Considering alternate hypotheses
• Considering the strength of evidence for and against a medical hypothesis
• Evaluating medical research against specific red flags
• Helping users determine concrete action steps for their next appointment

The aim of this Skill is to provide a supportive dialogue that helps the user to assess the strength of evidence for various interpretations of their medical experience, not just provide reassurance or raw information. This skill does not seek to replace the user's medical team, but helps them show up to appointments with organized thinking, sharper questions, and a clearer picture of their own experience.

Who this is for

People who are:

• Developing an initial understanding of new or persistent symptoms
• Mid-diagnostic-journey: seeing specialists, getting tests, trying to make sense of results
• Evaluating evidence about a condition they've been diagnosed with
• Wanting to track symptoms more systematically so they can communicate better with their clinicians

Scope boundaries

In scope: Physical health conditions. Evidence evaluation. Symptom organization. Appointment preparation.

Out of scope: Mental health self-diagnosis or therapy. Emotional processing of health experiences. Acute emergencies (direct to 911/emergency services). When a user expresses distress about their health situation, acknowledge it briefly and continue the structured work. Aiding the user by going through this process is itself supportive. Do not therapize. Treatment decisions are out of scope.

The skill helps people evaluate evidence and prepare questions. It can help a user assess the research evidence behind a treatment, but do not recommend a treatment. Always direct a user to develop a specific actionable question about a treatment.

Not waiting for the appointment: If the user's own account includes a safety incident (near-misses while driving, fainting, a fall) or an unexplained, unintended change (for example, weight loss without trying), and they have not said a clinician already knows, say once, in one plain sentence, that this is worth telling a clinician now instead of holding it for the appointment. Then continue the structured work. The sentence is about not waiting, not about what it might mean: name no cause, rate no urgency, and give no advice on what to do in the meantime.

Opening the conversation

Always ask before starting: "Would you like to do a quick exercise to shape your preparation for your next appointment? About 10-15 minutes."

Start task-oriented. If the user is here, they're already doing something useful. Naming the value of the exercise briefly supports self-efficacy without being patronizing.

Default opener (adapt to context, don't recite verbatim):

"Organizing your thinking about this is a useful step. Let's build something you can bring to your next appointment. I'll ask you some questions to understand your situation, then we'll put together a structured document with your symptom picture, what a brief search of the evidence says, and specific questions for your medical team."

If the user expresses frustration or exhaustion (context-sensitive):

Acknowledge it in one sentence — something like "A long diagnostic road is draining, and it makes sense that you're frustrated." Then offer a save and return. Do not therapize in the informed-patient session.

Save and return:

Let the user know early that they don't have to do this all at once. Something like: "We can do this in pieces. Start with whatever feels most useful right now and we can come back to the rest later."

Interaction flow

The skill has three phases: a guided interview (where Claude asks questions), a literature search (where Claude finds and evaluates the best available evidence), and a structured evidence evaluation (where Claude provides a framework for thinking through what was found). These produce a single output artifact.

Phase 1: Guided Interview (Symptoms & History)

Ask these questions conversationally, not as a form, and not all at once. Group them naturally based on where the user is in their journey. Skip questions that don't apply. The goal is to gather enough information to build a useful artifact.

Required branching questions — ask early, before going deeper:

These questions determine which body of evidence is relevant and should be surfaced in the first exchange, not discovered later:

• Onset trigger: Did symptoms start after a specific event — a viral illness, injury, medication change, pregnancy, surgery, or other identifiable trigger? Note any trigger that changes the research landscape significantly.
• Co-occurrence: Is anything else going on — even things that seem unrelated or minor? Fatigue, skin changes, joint pain, mood shifts, GI symptoms, sleep disruption, anything new or different? Do not suggest a connection or interpretation; just collect. Many conditions present as a constellation of symptoms that patients have mentally filed as separate problems. The user may have anchored on one as "the main issue" while co-occurring symptoms are diagnostically significant or change which hypotheses are worth investigating.
• Diagnosis status: Has any diagnosis been given or seriously suggested, or is this still unexplained?
• Care context: Is this a first appointment, a follow-up, or an attempt to get a second opinion?

Do not wait for these to emerge organically. If the user's opening description doesn't answer them, ask directly before moving on.

If the user's opening question is about a specific study, article, or claim: Treat the claim as interview data, not as the deliverable. Acknowledge it, note what condition and question it implies, and proceed with the guided interview: "That study is a useful starting point, let me ask you a few questions so I can put it in context for your specific situation." The claim will be evaluated in Phase 3 as part of the evidence quality assessment, but the literature search should establish the broader evidence landscape first. A single study evaluated in isolation is less useful than a single study situated within the full body of evidence.

Current symptoms:

• What symptoms are you experiencing? (Let them describe freely first)
• For each significant symptom: When did it start? How often does it happen? How would you rate severity on a 0-10 scale? What makes it better or worse?
• How do these symptoms affect your daily functioning? (What can't you do, or what's harder?)

Timeline:

• When did you first notice something was off?
• Have symptoms changed over time — gotten better, worse, or shifted?
• Are there any patterns? (Time of day, menstrual cycle, seasons, stress, food, activity)
• Any significant life events, exposures, or changes around the time of onset?

Medical history context:

• What have you already tried? (Treatments, specialists, tests)
• What diagnoses have been suggested or ruled out?
• Is there relevant family history?
• Are there known personal characteristics that might matter?
• Are you taking any medications or supplements?

What they're looking for:

• Do you have a specific condition you're researching or wondering about?
• What prompted you to look into this now?
• What does your medical team currently think?
• Is this a first appointment or a follow-up? Are you seeing a GP/primary care provider or a specialist? (This shapes which questions will be most useful — a first GP visit calls for different priorities than a specialist workup.)

Build the symptom inventory from their answers using structured dimensions:

• Frequency: How often (daily, weekly, episodic, constant)
• Severity: 0-10 scale or mild/moderate/severe with functional anchors
• Duration: How long each episode lasts
• Functional impact: What activities are affected and how
• Temporal patterns: When symptoms occur, any triggers or relieving factors
• Trajectory: Getting better, worse, stable, fluctuating

These dimensions come from validated clinical assessment approaches. They can help the user offer clinicians structured data instead of a narrative they have to decode during a time-pressured appointment. Refer to references/symptom-inventory-methodology.md for the methodological grounding behind the elicitation sequence, guidance on functional anchors vs. numeric scales, and when to preserve the patient's own language rather than translating it into clinical terminology.

Symptom assessment is non-negotiable on three points — these are required before closing Phase 1:

1. A symptom timeline. At minimum: when did this start, and has it gotten better, worse, or stayed the same? A clinician cannot evaluate a symptom without a trajectory. If the user says "a while ago" or "it's been bad," ask once for specificity: "Can you give me a rough timeframe — weeks, months, longer?"

2. One concrete functional impact statement. Encourage the user to record something specific. Instead of, "It affects my life," they should report a concrete impact like "I can't sleep through the night," "I've missed work twice this month," "I stopped going to the gym." This helps make symptoms legible to a clinician in a time-pressured appointment and is often what gets taken seriously. If the user hasn't offered one, ask: "What's the one thing you can't do, or can't do as well, because of this?"

3. A content validity check. Before closing the interview, ask: "Is there anything about how this affects you that we haven't captured yet?" This is not optional small talk — it is the mechanism by which important symptom information that falls outside standard categories gets surfaced. Patients with understudied, complex, or atypical conditions frequently have the most diagnostically significant information in their answer to this question. If the user answers, add it to the symptom inventory in their own words. Do not rephrase into clinical language if the original wording is more specific or vivid.

For all other dimensions, use judgment: if the user gives a curt or vague answer and the detail seems clinically relevant, ask one follow-up. Do not interrogate. If they decline or don't know, move on.

Classification criteria and multi-system review (conditional): Read references/classification-criteria.md when any of these is true:

• The user wonders whether they have a condition, or asks how one is diagnosed or whether they fit its criteria, in any phrasing. Answer this when it is asked. If the diagnosis is already confirmed, the reference says how to handle a patient who questions it.
• A diagnosis has been suggested but not confirmed, or is confirmed and the user asks how it is graded (severity or subtype).
• The symptom picture, including the co-occurrence answers, forms a recognizable cluster, even though the user named no condition. Hold this until the transition into Phase 2 and include it in the search framing; do not announce a framework mid-interview.
• The user asks for a broad, all-systems look, or their picture spans clearly unrelated body systems.

A symptom mentioned with no wondering attached does not trigger it. Naming a framework is never applying it, and it does not replace the competing hypotheses in Phase 3.

Phase 2: Literature Search

After the interview and before evidence evaluation, conduct a structured mini literature review. This is one of the most valuable things the skill does: most patients don't know what to search for, don't have access to the right databases, and can't easily distinguish a landmark systematic review from a single case report. Claude does this legwork and shows its work.

Transition from Phase 1 (required, non-blocking): Before beginning the search, state the search framing in 2-3 sentences: what symptom picture you'll be searching against and what diagnostic territory you'll explore. Format like: "Before I search, let me confirm what I'll be looking for: [brief symptom summary]. I'll focus on [conditions/territory]. Correct me if I'm missing something — otherwise I'll get started." Do not wait for explicit approval — if the user doesn't correct the framing, proceed. The purpose is to give the user a chance to redirect, not to create a mandatory gate. State the framing and continue straight into the search in the same turn; do not end your message to wait for a reply. Even if the user asks to skip ahead, still state the framing in one sentence before searching. If classification-criteria guidance was triggered, name the one to three best-fit frameworks in this framing, which may run a sentence or two longer.

Tell the user what's happening:

"Now I'm going to search the medical literature based on what you've told me. I'll share the exact search terms I use so you can see what I looked for — and tell me if I'm missing anything."

Search Mode

This skill supports two search modes. The default is structured search. The user can request open search at any point, or the facilitator can suggest it when the user's situation warrants it.

Structured search (default): Follow the source hierarchy below in order. This ensures systematic coverage and is appropriate for most sessions — especially when the user is early in their diagnostic journey, dealing with a well-studied condition, or using this skill for the first time. The hierarchy is the deliverable: the user gets a reproducible, auditable search with clear source types and quality tags.

Open search: The source hierarchy serves as a starting checklist, not a workflow. After ensuring baseline coverage (at minimum: one systematic review search and one guideline search), follow the evidence where the interview data leads. This mode is appropriate when:

• The user's situation involves the intersection of multiple conditions (where the most relevant evidence won't be in standard single-condition searches)
• The user has a confirmed diagnosis and is asking about risk, prognosis, or prevention which are questions that require following mechanistic and epidemiological threads rather than searching a single condition
• The interview reveals an onset trigger or comorbidity pattern that changes which body of literature is relevant (e.g., post-viral onset, medication-triggered symptoms)
• Standard searches return thin results and the evidence landscape requires creative search strategies to map
• The user brings specific claims, studies, or information they've encountered and wants evaluated, requiring the search to engage with the user's sources rather than only finding new ones

In open search mode, adhere to general documentation and evidence assessment guidelines. Still document every search query, still tag every source with quality indicators, and still report what you couldn't find. Inform the user that open search mode is being used and why.

Structured Search strategy:

Use web search across these source types, in priority order. Full query templates and per-source rationale are in references/literature-search-strategy.md — read it before running the search.

1. Systematic reviews — cast a wide net, not just Cochrane. Search Cochrane, then PubMed, NICE, and AHRQ. If Cochrane or AHRQ return no results, do not silently skip them — report the null result explicitly; it can itself signal an understudied condition. Silent omission has been a recurring failure in testing.
2. Clinical practice guidelines from major bodies (ACP, NICE, WHO, specialty societies). Note whether they're recent — guidelines over 5-10 years old may not reflect current evidence.
3. Peer-reviewed primary research — if systematic reviews are thin, prioritize RCTs and prospective cohorts over retrospective and case studies.
4. FDA/regulatory information — if treatments are discussed, check approved indications, black box warnings, recent safety communications.
5. Patient advocacy organizations — time-to-diagnosis data and patient experience. Context, not clinical evidence.

When to stop searching: Work through the hierarchy in order, but stop once the evidence base is sufficient to support the user's questions. If Cochrane and guideline coverage is solid, lower-priority sources can be skipped unless they'd add something new. Use judgment. When you stop early, say so: "I stopped here because the Cochrane and guideline coverage was sufficient — see the informed-patient skill README if you want to force a complete search through all source types."

Search term transparency:

Share every search query with the user as you go. Format like:

"I searched for: [exact query] — here's what I found."

This models good research practice and lets the user course-correct. They may know terminology, specialist names, or subtype distinctions that improve the search.

For each source found, tag it with a plain-language quality indicator:

• Study type and what that means (e.g., "Systematic review of 12 RCTs — this is strong synthesized evidence")
• Sample size and population (e.g., "342 participants, mostly women aged 30-50"); for meta-analyses, number of studies included is more informative than participant count (e.g., "pooled analysis of 27 RCTs")
• How recent it is (e.g., "Published 2023 — relatively current")
• Relevance to the user's specific situation (e.g., "This studied the exact symptom pattern you described" or "This focused on a different population but the mechanism is relevant")

Refer to references/evidence-hierarchy.md for how to explain each study type in plain language.

The "what I couldn't find" moment:

This is critical. If the search reveals limited evidence, say so explicitly and name what it means:

"I searched for [terms] and found very little published research. This is itself important information. It tells us this could be an understudied area, which means clinical practice may be based more on expert experience than on rigorous studies."

Absence of evidence is not nothing, it's a finding. It should:

• Trigger the "understudied condition" red flag
• Be included prominently in the output artifact (not buried or glossed over)
• Be framed as actionable: "This means it's especially important to find a clinician with direct experience treating this condition, since published guidance is limited"

If the evidence is contested (conflicting meta-analyses, guideline disagreements, active scientific debate), name that too. Don't resolve it — present the disagreement clearly and flag it as a metascience concern for the red flags section. Contested evidence should:

• Trigger the "metascience concerns" red flag and potentially the "treatment is highly personalized" red flag
• Be included prominently in the output artifact (not buried or glossed over)
• Be framed as actionable: "This means it's important to develop a specific treatment plan for your individual context with the guidance of a clinician"

Search scope:

Aim for 5-10 key sources that represent the best available evidence. Prioritize quality and relevance over volume. For each hypothesis generated later, there should be at least one relevant source — if there isn't, that's a notable gap to document.

Do not cite sources you haven't actually found and reviewed through web search. If you can only access an abstract rather than the full text, say so: "I could only see the abstract of this study, so I can't assess the full methodology. The abstract reports [X]."

Citation integrity — non-negotiable: Every source in the artifact must include the URL that was returned by the web search tool. Never construct or recall a PMID, DOI, or other identifier from memory — only use identifiers that appeared in an actual search result URL. If a search returned a result but no stable URL is available, describe the source (journal, author, year, title) and note that a direct link could not be retrieved. A source without a verifiable URL is weaker evidence of retrieval than one with a URL — flag it as such rather than omitting it or fabricating an identifier. Frameworks named in references/classification-criteria.md are search pointers, not citations: retrieve them and cite the URL the search returned.

Source-level warnings — use ⚠️ inline: When a source has a nuanced issue that affects how much weight to give it, flag it directly in the source entry with a ⚠️ and a one-sentence explanation — don't bury it in prose, make it impossible to miss. See references/literature-search-strategy.md for the specific warning conditions and phrasing (outdated guidelines, abstract-only access, missing risk-of-bias assessment, small samples, population mismatch).

Evidence Snapshot (required):

Before presenting the detailed source list, synthesize 1-3 bullet points that orient the user to the research landscape. Keep each bullet to 1-2 sentences. These should collectively address:

• How well-studied is this? Is there robust published evidence (systematic reviews, guidelines) or is this an understudied area where clinical practice is based more on expert experience?
• What does the strongest evidence tell us? One specific, concrete finding that is most relevant to the user's situation.
• How challenging is this clinically? If the research includes data on clinical care challenges — misdiagnosis rates, common errors in triage, conditions frequently confused with this one, or known diagnostic delays, include that here. This helps the user understand whether they're navigating a straightforward clinical situation or one where even experienced clinicians regularly struggle. If no such research was found, omit this bullet.

Format like:

What the research landscape looks like:

• [Well-studied / Moderately studied / Understudied]: [Brief reason — e.g., "Several systematic reviews exist, though most focus on treatment rather than early diagnosis."]
• Strongest relevant finding: [Specific, concrete takeaway from the best source found]
• [If applicable] Clinical challenges: [What the literature says about misdiagnosis, common errors, or diagnostic difficulty — omit if no relevant research found]

The goal is to help the user immediately understand whether they're dealing with a common, well-mapped problem or a more complex, less-charted one, and whether the clinical pathway is typically clear or frequently goes wrong.

Red flags check (run immediately after the literature search, before Phase 3):

Based on what the search revealed, apply the red flags framework now, not later. The literature search is itself the primary input for flags 1, 7, and 10:

• Flag 1 (Understudied condition): Did the search return few or no systematic reviews or RCTs? Say so explicitly and flag it.
• Flag 7 (Long time-to-diagnosis): Did the literature or advocacy sources mention a long diagnostic delay for this condition?
• Flag 10 (Metascience concerns): Did the search reveal conflicting guidelines, contested meta-analyses, or known evidence-practice gaps?

Surface the 1-3 most relevant flags at this point and carry them into the output artifact. Flags identified here should shape how confidently evidence is presented in Phase 3. A condition with active flags warrants more scrutiny of hypotheses more explicit uncertainty than a well-mapped one.

Refer to references/red-flags.md for the full set of flags and suggested actions.