Guide Updates
September 5, 2026
This revision connects the guide more closely to the Fritsche Lab and turns existing advice into practical starting points. The eight chapters, established page paths, sidebar, and search remain the core of the site.
- A lab-aligned reading experience: the lab’s FL² mark, navy and maize identity, teal accents, and responsive cards.
- The lab’s domain: canonical URLs, lab links, and citation metadata now use
fritschelab.org; the project path stays/practical-genai-coding-guide/. - A hands-on quick start: ten synthetic records, a scoped R task, manually specified expected results, and executable checks.
- A coding agent playbook: context selection, handoffs, bounded delegation, enforced permissions, and review, with a diagram showing the boundary between agent work and protected-data execution.
- Reusable task briefs and session handoffs alongside updated planning, refactoring, and documentation prompts.
- Research workflow updates: R dependency locking, pipeline dependencies, and separate evaluation when an LLM remains part of the running analysis.
- Explicit R setup: the simulator and R skeleton report missing packages instead of installing dependencies during execution.
- Corrections: BMI flooring arithmetic, unsupported demographic mappings in the advanced teaching plan, a nonexistent simulator flag, and prompts that could change output behavior or guess software versions.
Evidence memo
Current-source research and a focused methods audit preceded the content edits. The table records the selected changes and why they matter. “High confidence” refers to the documented behavior; examples and workflow combinations are editorial adaptations for this guide.
| Update target | Evidence and rationale | Priority / confidence |
|---|---|---|
| Chapter 3, playbook, task brief | Official coding guidance and instruction discovery support explicit tasks, relevant context, and verification commands. | High / high |
| Handoff template and playbook | Execution-plan guidance supports persistent decisions, progress, and next steps; the handoff template is our adaptation. | High / high |
| Chapter 4 and playbook | Agent security documentation distinguishes instructions from controls and explains the limits of shell network restrictions. | High / high |
| Quick start and Chapter 6 | Testing expectations support precise result checks; the synthetic cases encode this guide’s explicit task contract. | High / high |
| Chapters 7–8 and documentation template | renv supports package restoration, while targets tracks declared analysis dependencies. Neither validates the scientific method. | High / high |
| Playbook and tool appendix | ellmer documents optional fields; evaluation guidance supports task-specific datasets and repeat evaluation. | High / high |
| Documentation prompt | Miniforge documents the retirement of Mambaforge. Existing project manifests and tested versions take precedence over guessed setup instructions. | High / high |
| BMI teaching plans and Chapters 1–2 | NICE NG246 does not establish the old mapping from broad EHR race labels to its specified population groups. The tutorial now uses an explicit synthetic cohort and general descriptive categories. | High / high for the mapping correction |
The new workflow sources were checked on September 5, 2026 and are listed in References. NICE’s official indexed content was available, while direct automated requests to the chapter/PDF returned access restrictions. The audit corrected the source-to-data mismatch without extending clinical recommendations. The tutorial’s age-at-measurement restriction is a teaching choice, not an attributed NICE age cutoff.
This revision does not revalidate every historical citation. It avoids model rankings, price comparisons, guaranteed productivity claims, and tool-specific settings that would age quickly. Source dates identify what was checked; they do not guarantee that a provider’s documentation will remain unchanged.
Suggest an improvement
Open an issue with the page, a concrete correction or use case, and a primary source where appropriate. For a runnable example, include a small synthetic input, the expected result, and the command you tried. The development guide explains how to preview changes locally.