Grounded in real guidelines, gated for safety, and set up inside Claude in about ten minutes. No coding. You copy an instruction block, paste it into a Claude project, and start.
01
Evidence
Search & appraise papers
Ready
02
Teach & Communicate
Decks & patient letters
Ready
03
Clinical Reasoning
DDx · options · MDT · follow-up
Setup
04
Rx Cross-Check
Prescription safety check
Pharmacist
Three rules that never bend
You decide, always. Every output is a draft for your review — never an order, never a diagnosis.
No patient identifiers. Remove names, IDs and dates before you paste any case. The assistants will refuse identifiable input.
If it isn't sourced, it says so. No dose or citation is invented. “Not in my knowledge base” is a correct, safe answer.
Set up in about 10 minutes
Five steps, once.
Do this for the Evidence assistant first — it works right away and touches no patient data. Add the others when you're ready.
Open Claude and create a project
Go to claude.ai → Projects → Create project. Name it ONCO-EVIDENCE.
Copy the instructions
On the Evidence card below, press Copy instructions. In your project, open Set project instructions and paste.
Add the starter files
Download the files listed on the card and upload them to Project knowledge. For Evidence this is quick — the safety files and a couple of references.
Test it before you trust it
Ask it for a drug dose. It should refuse and tell you to verify — not answer from memory. If it confidently gives a dose, re-check that HARD_GATES.md uploaded.
Use it
Ask a real literature question from your week. When it feels natural, come back and add Teach & Communicate, then the clinical assistants with your pharmacist.
The four assistants
Copy, paste, done.
Each assistant is one Claude project. The instruction block is what makes it safe — copy it in full. Build them top to bottom; the order is deliberate.
ONCO-EVIDENCE
Ready
Evidence
Search, appraise and synthesise the literature — with every citation verified.
Start here · works right away
The safest place to begin. It touches no patient data and needs almost no setup — paste the instructions and ask it a literature question.
Only build this after the Tier B local files (formulary, capability map) are filled — otherwise it will suggest tests your centre cannot run. Run it in shadow mode for 60 days before it advises anyone.
Upload to the project:
all control filesA1 guideline PDFA2_STAGINGA3_LANDMARK_TRIALSA4_REGIMEN_SHEETSall Tier B filesC1_NOTE_TEMPLATESC2_MDT_PROFORMA
ONCO-RX-GUARD
Pharmacist required
Prescription Cross-Check
An independent second check on a chemotherapy prescription: doses, adjustments, interactions, omissions.
Pharmacist sign-off required before clinical use
The highest-stakes assistant. It must not go live until your clinical pharmacist has filled and signed the dose and interaction files, and until it passes all twelve red-team tests. Until then it answers every dose question with “not in my knowledge base” — which is correct, not broken.
Upload to the project:
all control filesA4_REGIMEN_SHEETSA5_DOSE_ADJUSTMENTA6_INTERACTIONS_COREB1_FORMULARYB2_LOCAL_PROTOCOLS
The knowledge base · for you and your pharmacist
The clinical files you fill in.
These are empty templates on purpose. The doses, interactions and formulary are not filled in — because an assistant that invents them is worse than no assistant at all. You and your clinical pharmacist fill them from NCCN / BNF / your local protocol. Files marked sign-off must be signed before the Rx Cross-Check assistant may be used on a patient.
Control · load into every project
HARD_GATEScontrol
The eight safety rules. Loaded into every project.
SOURCE_REGISTRYcontrol
Provenance table + the pharmacist sign-off ledger.
GLOSSARYcontrol
Local brand↔generic map and unit conventions.
REDACTIONcontrol
De-identification checklist. Run before any case.
CHANGELOGcontrol
Version log. A KB change is a code change.
Tier A · Authority (what is true)
A1_GUIDELINES_READMEAuthority
How to add guideline PDFs (one site at a time).
A2_STAGINGAuthority
AJCC / WHO staging tables — you paste these in.
A3_LANDMARK_TRIALSAuthority
Verified trial registry — the anti-hallucination anchor.
A4_REGIMEN_SHEETSAuthority · sign-off
Regimens, doses, ceilings. Pharmacist-owned.
A5_DOSE_ADJUSTMENTAuthority · sign-off
Renal / hepatic / haem bands + the formulae. Pharmacist-owned.
A6_INTERACTIONS_COREAuthority · sign-off
Interactions, with mechanisms. Pharmacist-owned.
Tier B · Local (what is possible here)
B1_FORMULARYLocal · sign-off
What your pharmacy actually stocks and its cost.
B2_LOCAL_PROTOCOLSLocal
Your unit's FN, extravasation, TLS pathways.
B3_PATHWAYSLocal
Real waiting times and referral routes.
B4_CAPABILITY_MAPLocal
What your centre can and cannot do.
B5_COST_TIERSLocal
Out-of-pocket cost per regimen.
Tier C · Form (how output should look)
C1_NOTE_TEMPLATESForm
Clerking, on-treatment, toxicity note templates.
C2_MDT_PROFORMAForm
One-page tumour board proforma.
C3_SLIDE_SKELETONSForm
Journal club and case-presentation skeletons.
C4_PATIENT_COMMSForm
Tone rules for bilingual patient letters.
C5_PROMPT_LIBRARYForm
Ready-made prompts for all four projects.
Prefer everything at once?
The full safety charter — the eight rules behind every assistant
Physician-in-the-loop, always. Outputs are drafts ending in a sign-off line. Never orders.
Source-grounded only. Every dose, threshold, interaction and citation traces to a file.
The not-in-knowledge-base gate. If a fact isn't in the files, the assistant says so rather than guess.
Zero fabricated citations. Every reference carries a real PMID or DOI, or it is deleted.
No patient identifiers. De-identify first; the assistant refuses identifiable input.
Not a medical device. Educational and administrative support — it does not diagnose or prescribe.
Guidelines are the ground truth. The AI is a faster route to them, never a replacement.
Audit trail. Every clinical run is logged: inputs, knowledge-base version, output, who signed.