Research, Knowledge Management & Clinical Content Creation

The Diagnostic Prompt^IMI^ Revisited + The Convergence Check^IMI^ — Hallucination, Sycophancy and Reproducibility

Opens with a formal revisit of The Diagnostic Prompt^IMI^, now applied to research, teaching and audit questions: before you open any search tool, let AI interrogate the question. Names the three prompt-engineering techniques participants already used in IMI-01 — few-shot (custom scribe templates), chain-of-thought and negative prompting (Claude Skills) — and maps each to its industry term. The remainder introduces three measured failure modes of large language models — hallucination, sycophancy and non-reproducibility — with live demonstration of each, and introduces the Convergence Check^IMI^: a four-move serial-testing framework (Reframe, Repeat, Widen, Report) that catches drift toward the user and drift between runs structurally. Live comparison across Claude, ChatGPT and Gemini. Verified evidence is stored in Gemini Notebook as a grounded knowledge base, foreshadowing Week 2.

Reflection on Advanced Prompting — Reviewing AI Tool Use Practice

Participant applies The Diagnostic Prompt IMI to one real complex clinical or research question from their own practice — a question where standard prompting has previously given insufficient results. They then use advanced prompting techniques (chain-of-thought, few-shot, or negative) to investigate it across two platforms and compare outputs. Structured reflection on their own AI tool utilisation practice.

AI for Clinical Data, Part 1 — Claude, Copilot & GPT in Excel

Structured online teaching session built around a shared 25,000-row simulated ED dataset, used identically across all three tools so outputs are directly comparable. Covers: orienting each AI tool to a dataset before asking analytical questions — describing structure, column names, and value formats in a single upfront prompt; the golden rule of data prompting — name the exact column, specify the output format, state the comparison; a live side-by-side demonstration across Claude, Copilot, and GPT on doctor-level MET call rates, missed-opportunity rates, shift patterns, and boarding-delay outcomes; an honest comparison of the three tools (Claude for ambiguous or multi-step reasoning, Copilot for native spreadsheet integration, GPT for accessibility without spreadsheet-native tools); a first pass at auditing AI-generated findings — checking one figure against a known reference value before accepting it; a brief governance note on identifiable staff data in simulated datasets, and how the same Australian Privacy Principles distinction applies to real hospital data.

AI for Clinical Data, Part 1 — Claude, Copilot & GPT in Excel

Participant works directly with the MET call dataset (or their own real/de-identified audit data, where ready) using one or more of the three tools, running at least three questions from the Tier 1–2 question bank — doctor performance, boarding delay, shift and timing, clinical patterns. For at least one finding, participant applies the Convergence Check^IMI^ — cross-checking the AI's figure against a documented reference value, a second tool, or manual recalculation — before treating it as reliable. Structured reflection on data analysis practice and AI output verification.

AI for Clinical Data, Part 2 & Presentations — Governance Reporting with Claude, Copilot & GPT

Structured online teaching session continuing directly from Week 2, using the same dataset and three-tool comparison. Data governance: advanced questions from the Tier 3–4 bank — missed-opportunity and early-warning failure analysis, quarterly and trend reporting, pattern recognition across combined variables; asking AI to write a governance-ready paragraph or board summary from the findings, then applying the Convergence Check^IMI^ and a sycophancy check to that summary — does every figure still hold up, has it overstated certainty anywhere — before it could go to a real Patient Safety Committee. Presentations (kept brief, given limited teaching need): briefing Claude or Copilot to turn an audited governance paragraph and chart into slide content for a grand rounds, M&M, or governance committee presentation; a quick practical decision framework for choosing the right tool and environment.

AI for Clinical Data, Part 2 & Presentations — Governance Reporting with Claude, Copilot & GPT

Participant selects one or more Tier 3–4 questions and produces a written governance-style summary from the MET call dataset (or their own audit data), applying the Convergence Check^IMI^ and a sycophancy check to their own AI-generated summary before finalising it, then briefs Claude or Copilot to turn that audited summary into a short slide set. Participant documents what changed between the AI's first draft and the audited final version. Structured reflection on the combined data-audit-to-presentation workflow.

Complex Skills and Working with Real Data

Working live session. Participants arrive with a more complex Claude or ChatGPT skill and a dataset — a practice spreadsheet emailed that week, or their own de-identified audit or roster file. Session time is used to stretch the skill against real-shaped data and check where a human still has to intervene.

Reflection — Complex Skill, Data and What You Learned

Participant writes up what they learned from building a more complex skill and running it against a real-shaped dataset — including where it broke and what they would not trust it to do alone.