ExploringGenerativeAI
title: Exploring Generative AI (Fowler Memos) type: technique created: 2026-05-21 last_updated: 2026-07-11 related: ["radar/techniques/AIConversationTypes", "radar/techniques/StructuredPromptDrivenDevelopment", "radar/techniques/LLMTDDLoop"] sources: ["https://martinfowler.com/articles/exploring-gen-ai.html", "https://martinfowler.com/articles/exploring-gen-ai/local-models-for-coding-factors.html"] radar_quadrant: Techniques radar_ring: Assess radar_position: inner
A running series of practitioner memos on martinfowler.com documenting observations, patterns, and cautions around integrating generative AI into software development workflows, written by members of the Thoughtworks technology advisory board.
The series uses a memo format — short, dated observations from practitioners actively using LLMs — rather than comprehensive articles. Each memo captures a specific insight: a failure mode observed in production, a prompt pattern that worked, a limitation encountered, or a workflow integration that changed how work gets done.
Topics covered across the memo series (as of late 2023):
- Prompt engineering patterns for code generation, test writing, and documentation.
- LLM limitations in reasoning tasks and code correctness — what models get systematically wrong.
- Pair programming with LLMs — when the AI is useful as a pair and when it introduces noise.
- Context window constraints — how limited context affects multi-file refactoring tasks.
- Testing LLM-generated code — the gap between code that passes tests and code that is correct.
- Retrieval-Augmented Generation (RAG) for private codebases — patterns and trade-offs.
Martin Fowler's site has a track record of publishing practitioner-validated patterns (Refactoring, Patterns of Enterprise Application Architecture, Continuous Integration). The generative AI memo series applies the same practitioner rigour to an area where hype significantly outpaces grounded experience. The memos are conservative: they note where LLMs fail as carefully as where they succeed.
A later memo in the series assesses whether LLMs run entirely on a developer's own machine, instead of a cloud service, are viable for coding work. The conclusion is not yet: local models are not a "plug and play" experience for developers. The memo identifies the working memory needed to hold both the model and a large enough conversation history (context window) as the main bottleneck, and names Qwen3.6 35B, a mixture-of-experts model, as the current best compromise between capability, speed, and memory footprint on well-specced hardware. Even with that model, output quality is inconsistent and reliably invoking coding tools (reading and editing files) remains shaky, partly because different local-model hosting tools disagree on how they expect a model to call those tools.
The memo series contextualises several blips already on the radar:
- radar/techniques/LLMTDDLoop — the memo on testing LLM-generated code addresses the same failure modes.
- radar/techniques/StructuredPromptDrivenDevelopment — the REASONS canvas formalises prompt discipline the memos describe informally.
- radar/techniques/AIConversationTypes — the NNG taxonomy and the Fowler memos are complementary frameworks for the same problem space.
Exploring Generative AI sits in the Assess ring of the Techniques quadrant, at inner position. First studied via martinfowler.com memo #8 (2023-11-30). The memo series is a living document; new entries appear as practitioners accumulate production experience. Inner position reflects high signal quality from a credible platform and direct relevance to any team integrating LLMs into a software development workflow. No tooling required — the value is calibrated practitioner perspective on where to invest and where to be cautious.
