Drug Pipeline Research Analyst

A folder-based AI researcher that investigates the drug development pipeline for a disease. Drop this folder into a Claude project and you have a research partner that maps trials, regulatory decisions, and the credibility of claims — then audits its own work with a hostile reviewer before handing anything back.

Default demo disease: Alzheimer's. Multi-disease use is supported.

---

How to use it

Setup (do this once)

1. Create a Claude project - can use an existing one
2. Copy the contents of DrugResearcher folder to your project folder including all subfolders
3. Enable web search in the project if not already
4. Paste the block below into the project's Custom Instructions field. This is what reliably routes user questions into the procedures rather than letting Claude default to summarizer behavior.

You are the Drug Pipeline Research Analyst. Drugs only — no devices,
supplements, lifestyle interventions.

BEFORE answering any drug/disease question, ASK three things in one short
message: (1) disease, (2) angle (landscape / one drug / regulatory tempo /
mechanism debate / failure pattern), (3) what they already know. Do not
produce findings, summaries, or lists until those are answered.

THEN route to a procedure and execute it verbatim:
- "what's in the pipeline / what's new / approved / scan" → procedures/scan.md
- "tell me about [drug] / deep dive on [drug]"            → procedures/deep_dive.md
- "audit / verify / fact check / are these real"          → procedures/source_check.md
  (execute as "The Auditor" — hostile reviewer; a clean audit with zero
   findings is a failed audit, not a successful one)

OUTPUT FORMAT — MANDATORY for scan and deep_dive:
- MUST render as a Claude Artifact (text/markdown). Title it [disease]-scan-YYYY-MM-DD
  or [disease]-deepdive-[drug]-YYYY-MM-DD. Do NOT render inline in chat.
- 2-4 sentence Overview synthesis at top (pattern, not list)
- Each drug entry: Mechanism / Status / Confidence (high/med/low)
- Each source row: **Name** — Tier [1-7], YYYY-MM-DD / link /
  [ ] link verified  [ ] claim matches pre-registration
- "What's missing" section per drug (non-empty)
- "Flags" section per drug
- Closing line: "⚠️ Unverified. Run source_check before trusting this."

Source tiers (rules.md Rule 2):
1 FDA/EMA primary docs · 2 ClinicalTrials.gov · 3 NEJM/Lancet/JAMA
4 Conference abstracts · 5 STAT/Endpoints journalism · 6 Press releases
7 Forums/advocacy

A freeform paragraph response to a landscape or drug question is a contract
violation. If you start drafting one, stop and run the procedure instead.

---

Using the researcher

1. For looking into drugs for a disease:
   • Scan is the keyword in the prompt: "Run the scan on Alzheimer's." or "Scan the Alzheimer's drug pipeline."
2. For information on a single drug:
   • Deep dive is the keyword "Run a deep_dive on donanemab." or *"Deep dive: lecanemab."
3. Running a link self verification
   • Source check is the keyword "Audit the scan." or "Run source_check on the donanemab dossier.

Natural-language prompts also work but are less reliable. The custom instructions block above is what makes them route correctly:

• "What's in the Alzheimer's drug pipeline right now?" → should trigger scan
• "Tell me everything on donanemab." → should trigger deep_dive
• "Are these sources real?" → should trigger source_check

After any scan or deep_dive, run source_check before trusting the output. The artifact will end with a ⚠️ unverified handoff reminding you.

What "verified" means

Every source the researcher cites comes with two empty checkboxes:

- **CLARITY-AD NEJM publication** — Tier 3, 2023-01-05
  - What it tells us: Lecanemab slowed decline by 27% on CDR-SB
  - Link: https://www.nejm.org/doi/10.1056/NEJMoa2212948
  - [ ] link verified  [ ] claim matches pre-registration

The boxes are deliberately empty when the scan or deep_dive first runs. They're a promise that the researcher has not yet checked them. They get filled in only when you run source_check, which is a separate, adversarial pass that re-opens every source.

• [x] link verified — The Auditor opened the link and confirmed it goes to the document the researcher claimed it does (not a press release dressed up as a peer-reviewed paper, not a 404, not a different paper with a similar title).
• [x] claim matches pre-registration — For trial results only: The Auditor pulled the trial's original registration on ClinicalTrials.gov and confirmed the endpoint the publication reports is the same endpoint the trial was designed to test. If the trial changed its endpoint mid-study (a credibility red flag called "outcome switching"), this box stays empty even if the link is fine.

After source_check runs, a passing source row looks like:

- **CLARITY-AD NEJM publication** — Tier 3, 2023-01-05
  - What it tells us: Lecanemab slowed decline by 27% on CDR-SB
  - Link: https://www.nejm.org/doi/10.1056/NEJMoa2212948
  - [x] link verified  [x] claim matches pre-registration

A failed verification leaves the box empty and adds a one-line note explaining what's wrong:

- **Biogen Q1 press release** — Tier 3, 2024-03-15
  - What it tells us: Phase 3 trial "successful"
  - Link: https://investors.biogen.com/...
  - [x] link verified  [ ] claim matches pre-registration
  - NOTE: Tier was listed as 3 but this is a press release — corrected to Tier 6. Press releases cannot be checked against pre-registration.

If you see unchecked boxes on a scan you've never audited, that's expected. If you see unchecked boxes after source_check ran and there's no NOTE explaining why, ask the researcher to re-run the audit — that's a gap in the verification record.

Switching diseases

The methodology in rules.md and procedures/ is disease-agnostic. To switch:

1. State the disease at the start of the session.
2. The researcher will load reference/credibility_flags.md (cross-disease) for methodology.
3. The Alzheimer's-specific reference/alzheimers.md will not apply. If you want disease-specific landmines for another disease, ask the researcher to draft one — that file is the only disease-specific piece of the folder.

---

What this researcher does

• Maps the active clinical trial landscape for a disease's drug pipeline.
• Tracks FDA and EMA approval status, the pathway used, and any advisory committee controversies.
• Evaluates source credibility and flags what's paywalled, unpublished, or missing from the public record.
• Synthesizes patterns across the pipeline rather than listing trials.
• Asks what angle you're working before producing anything.
• Runs an adversarial verification pass against its own scans and deep dives.

What this researcher does NOT do

• Cover non-drug interventions: devices, deep brain stimulation, digital therapeutics, cognitive training, supplements, diet, lifestyle interventions, surgery. (See identity.md for the full out-of-scope list.)
• Summarize the first article it finds.
• Treat pharma press releases as evidence.
• Present statistical significance as proof of clinical benefit.
• Give medical advice or make treatment recommendations.

---

Terms in the results

Plain-language definitions of terms that will show up in a scan or deep_dive artifact. Skip this section if you already know the field.

Phase 1 / 2 / 3 — The three stages of human drug testing. Phase 1 is small and asks "is it safe?" Phase 2 is larger and asks "does it appear to work?" Phase 3 is the largest and asks "does it work well enough in the real target population to justify the risks?" The majority of drugs that pass Phase 2 still fail Phase 3.

Approved / Pending / CRL / Withdrawn — Regulatory statuses. Approved means the FDA or EMA has cleared the drug for sale. Pending means the application is under review. CRL (Complete Response Letter) means the FDA reviewed the application and is asking for more information or rejecting it as submitted — not the same as a denial, but not approval either. Withdrawn means the sponsor pulled the drug from the market (sometimes voluntarily, sometimes under pressure).

Accelerated Approval — A faster FDA pathway that lets a drug reach the market based on a "surrogate" measurement (like brain plaque reduction) before clinical benefit has been confirmed. The sponsor is required to run a confirmatory trial afterward to prove the drug actually helps patients. Sometimes those confirmatory trials report years late, or never. Aduhelm in 2021 is the canonical example of accelerated approval going sideways.

Surrogate endpoint — A measurement that stands in for the clinical outcome that actually matters. Plaque clearance in the brain is a surrogate for whether memory and cognition actually improve. Tumor shrinkage is a surrogate for living longer. Surrogates are useful, but they don't always predict the real outcome — that's a recurring credibility issue this researcher will flag.

Advisory Committee (AdComm) — An independent panel of outside experts the FDA convenes to vote on whether a drug should be approved. The FDA isn't required to follow their vote, but historically has most of the time. When the FDA approves a drug over its own advisory committee's objection — as it did with Aduhelm in 2021 — that's a significant credibility flag worth surfacing.

ARIA (Amyloid-Related Imaging Abnormalities) — A safety signal specific to anti-amyloid Alzheimer's drugs (lecanemab, donanemab, and the now-withdrawn Aduhelm). It shows up as brain swelling (ARIA-E) or microbleeds (ARIA-H) on MRI scans. Most cases are detected on routine monitoring and have no symptoms, but it can be serious and occasionally fatal — especially in patients with a particular gene variant (APOE ε4) or on blood thinners.

Outcome switching — When a clinical trial reports a different "main result" in its publication than it originally promised to test in its registration. This is a credibility red flag — pre-registration exists specifically to prevent it — and the researcher will name it explicitly when it appears.

Pre-registration — Before a clinical trial begins, the sponsor is supposed to publicly register it (on ClinicalTrials.gov in the US) with the trial's primary question and how it will be measured. This creates a timestamped record so changes after the fact can be caught.

Source Tiers (1–7) — The researcher's reliability hierarchy. Tier 1 is primary regulatory documents (FDA review packages, EMA assessment reports). Tier 2 is trial registrations themselves. Tier 3 is peer-reviewed publications in major journals (NEJM, Lancet, JAMA). Tiers 4 and 5 are conference abstracts and science journalism — useful for leads, weaker for claims. Tiers 6 and 7 are press releases and forum/advocacy posts — treat as advocacy, not evidence. Every source in an artifact is labeled with its tier; never assume tiers are equivalent.

Confidence: high / medium / low — The researcher's honest read on how much weight to put on a given claim, based on the tier and quality of sources behind it. A claim backed by an FDA review document and a peer-reviewed pivotal trial is high confidence. A claim that rests on a press release is low.

"What's missing" — A required section in every drug entry. Names the gaps in what's publicly known: unposted trial results, paywalled datasets, populations excluded from trials, post-market commitments that haven't been honored. This is where the researcher does its most important work — surfacing absence is what separates investigation from summarizing.

Flags — Specific credibility patterns the researcher names when they apply: outcome switching, surrogate-endpoint claims presented as clinical benefit, accelerated approval with overdue confirmatory trial, industry funding without independent replication, and others. The full list lives in reference/credibility_flags.md.

---

The pipeline of procedures

The researcher's substantive work happens in three named procedures, each producing its own artifact:

Procedure	What it does	Artifact
procedures/scan.md	Disease-wide map: every drug in active development, plus any with regulatory action in the last 3 months. Synthesis first; drug list supports it.	[disease]-scan-[YYYY-MM-DD].md
procedures/deep_dive.md	Full dossier on one drug: trial history, regulatory record, safety, funding, contested questions, what's missing.	[disease]-deepdive-[drug]-[YYYY-MM-DD].md
procedures/source_check.md	The Auditor — a hostile, adversarial pass that re-fetches every source and confirms each claim. Updates checkboxes in the target artifact and appends to a cumulative check_log.md. A clean audit with zero findings is treated as a failed audit.	edits the target artifact + appends to check_log.md

The asking-before-answering rule (rules.md Rule 1) runs before any procedure. Even if you say "run the scan," the researcher confirms the disease and angle in one short message first.

If the researcher drifts into summarizer mode

If you ask a landscape question and get back a freeform paragraph with bullet-listed sources instead of a structured scan artifact — the procedure didn't run. Respond with: "That's not the scan procedure. Re-run as procedures/scan.md and produce the named artifact." The researcher should restart and execute the procedure correctly.

---

Folder contents

File	Purpose
identity.md	Who the researcher is, what's in scope, what's explicitly out of scope, limitations disclosed up front
rules.md	Investigative methodology — source hierarchy, asking-first, surfacing absence, uncertainty levels
examples.md	Side-by-side comparisons of summarizer behavior vs. investigative behavior, plus a worked pipeline example
procedures/scan.md	The disease-wide pipeline scan procedure
procedures/deep_dive.md	The per-drug deep-dive procedure
procedures/source_check.md	The adversarial source audit procedure (The Auditor)
reference/sources.md	Tiered source list with links to authoritative repositories
reference/pipeline_stages.md	Trial phases, FDA/EMA approval pathways, key terminology
reference/alzheimers.md	Alzheimer's-specific landmines: amyloid hypothesis debate, Aduhelm precedent, ARIA, key sponsors
reference/credibility_flags.md	Cross-disease credibility patterns: outcome switching, surrogate endpoints, ghost authorship, etc.
check_log.md	Cumulative source audit log — appended by every source_check run
refinement.md	Internal scratchpad — decisions, output format, pipeline design notes

---

Why an investigative researcher, not a summarizer

A summarizer reports what's in front of it. A researcher asks what's missing.

In drug development, the gap between what companies publish and what the full record shows is where the most important information lives — failed trials that quietly disappeared, accelerated approvals where the confirmatory trial never reported, advisory committees overruled by the agency they advise. None of that surfaces in a summary. It only surfaces when you know where to look and what to ask.

This researcher does three things most "research assistants" don't:

1. It asks before answering. No summary until the angle is established.
2. It runs a real pipeline. Scan → deep_dive → audit. Each step produces a discrete, verifiable artifact.
3. It audits its own work with a hostile reviewer. The Auditor exists specifically to catch what the upstream work got wrong. A clean audit with no findings is treated as suspicious, not as success.

That's the assignment.