Your scientists have ideas for tools they need — assay analysis pipelines, protocol trackers, molecular property calculators, screening dashboards. They shouldn't have to wait quarters for engineering to build them. AI-assisted development turns those ideas into production-grade software in days, with an auditable quality trail.
Every biotech R&D team generates more ideas for custom software tools than they can build. The computational chemist needs a molecular property calculator. The assay development team needs a protocol version tracker. The screening group needs a plate reader analysis pipeline. The genomics team needs a variant annotation tool.
If you have dedicated software engineers, they're backlogged. If you don't — and most small biotechs don't — the tools either don't get built, or they get built as fragile scripts that nobody tests, nobody validates, and nobody can maintain when the scientist leaves.
AI coding agents can build a working prototype in hours. But left unconstrained, the AI writes tests that confirm its own implementation rather than independently specifying behaviour. A defect in the reasoning produces a matching defect in the tests. Everything passes. The bug ships.
Real example: During hardening of a cell confluency assessment tool, a boundary condition test discovered that mask[-0:] selects an entire NumPy array instead of an empty slice — causing incorrect image processing output in assay analysis. It was caught only because the test was written before the implementation existed. Without that constraint, this bug silently corrupts results that inform drug candidate selection.
In your domain, a software quality bug is a scientific integrity bug. Speed without quality is not acceptable. You need both.
The methodology separates the work into three tracks. Phase 1 has two paths depending on the type of application: interactive apps (dashboards, GUIs) built with Cursor AI, or non-interactive apps (pipelines, CLIs, APIs) built with Claude Code. Both use the Claude app for strategic planning. Phase 2 then hardens any prototype into production-grade code under strict enforcement, also using Claude Code with the Claude app.
Phase 1 (Interactive) uses Cursor AI with the Claude app: the scientist describes a workflow in domain language, Cursor implements visual tools with live previews. Working prototype in hours. Phase 1 (Non-interactive) uses Claude Code with the Claude app: the same rapid prototyping, but for pipelines, CLIs, and data processing tools that don't need a GUI. Phase 2 uses Claude Code with the Claude app to productize: test-first development enforced by filesystem locks and SHA-256 hash audits, with a 30-check independent production readiness assessment. The handoff artefact is running code, not a document.
The methodology works for any R&D tool your team needs — whether it touches physical instruments in the lab or runs computations on molecular data. The same lifecycle applies: prototype with Phase 1 (interactive or non-interactive), then productize with Phase 2.
Version every protocol change. Link results to exact versions. Full audit trail for regulatory.
Upload raw data, fit dose-response curves, calculate IC50 values with confidence intervals.
Track Z-prime, signal-to-noise, and control drift across plates and campaigns.
Parse output from readers, liquid handlers, and imagers into structured databases.
Compute ADMET properties, Lipinski descriptors, and custom scoring functions via API.
Dock compound libraries, rank by score, filter by property criteria, export hit lists.
Annotate variants, map to pathways, score functional impact, generate reports.
Extract relationships from papers and transcripts, build queryable networks.
Teal = lab workflows. Purple = in-silico workflows. Both follow the same three-track lifecycle. Scope profiles adapt to the domain.
Biotech R&D teams naturally contain both halves of this workflow. The scientist who needs the tool and the data scientist or developer who could productize it are often in the same group — sometimes the same person. The handoff between Phase 1 and Phase 2 happens within your team, not across an organisational boundary.
Describe the tool you need in domain language. The AI builds a working prototype. You test it on real data. Stop waiting for the engineering queue.
Review architecture and test quality during hardening. Your skills are used for judgement, not implementation. Every commit is atomic and traceable.
Tools built in days, not quarters. Auditable quality trail. Production-assessed code your QA team can evaluate.
A 30-person biotech with 2 computational scientists and no dedicated software engineers is currently limited by how many tools those 2 people can build and maintain. If each of them can produce production-assessed tools at 5–10x the rate, that is not a marginal improvement — it is a structural change in what the team can attempt.
The three tracks are documented in detail on their dedicated sites.
Scientists build working dashboards, GUIs, and visual tools using Cursor AI with the Claude app. Setup, workflow, practice exercises, and honest assessment of capabilities and limits.
Scientists build pipelines, CLIs, APIs, and data processing tools using Claude Code with the Claude app. Same rapid prototyping, different tooling for headless workflows.
Claude Code with the Claude app productizes any Phase 1 prototype. VP-model orchestrator, enforcement mechanisms, scope profiles, and 30-check production readiness assessment.
Dr Raminderpal Singh · raminderpal@hitchhikersai.org · raminderpalsingh.com
