Exploring Generative AI
by Birgitta Bockeler (ed.), Martin Fowler · martinfowler.com / Thoughtworks
An ongoing memo series from Thoughtworks engineers, running since 2023 and still updated, examining AI coding assistants through practitioner experiments: context engineering, spec-driven development, TDD inside agent loops, supply-chain risk, and where the tools actually fail. The most rigorous longitudinal record of AI-assisted engineering anywhere.
More resources on AI-Assisted Coding
Beyond Vibe Coding: From Coder to AI-Era Developer
O'Reilly title on moving past prompt-and-hope workflows toward disciplined AI-assisted engineering: specification, verification, code review of generated output, managing technical debt from AI code, and how senior judgment changes when the model writes the first draft. The only book-length treatment of this topic that argues against the hype rather than selling it — its thesis is that vibe coding is a prototyping mode and everything past a prototype needs specification and verification.
DORA State of AI-assisted Software Development 2025
Google's DORA research team surveyed nearly 5,000 technology professionals on AI adoption in software delivery. Finds AI amplifies existing team capability rather than creating it, and identifies seven organizational capabilities that determine whether adoption helps or hurts.
Do Users Write More Insecure Code with AI Assistants?
First large-scale user study of people solving security-sensitive programming tasks with and without an AI assistant. Those given assistance wrote significantly less secure code across several languages, yet were more confident their solutions were secure. The finding that survives model upgrades is behavioural, not model-specific: assistance induces misplaced confidence, and participants who interrogated and re-prompted the tool produced safer code.
Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
Randomized controlled trial in which sixteen experienced open-source maintainers completed 246 real issues from their own repositories. AI access increased completion time by 19 percent, while the same developers believed it had sped them up.
Here's How I Use LLMs to Help Me Write Code
Willison's account of his day-to-day LLM coding workflow, covering realistic expectations, training cutoffs, context control, asking for options, testing everything the model writes, and when the human should take the keyboard back. Fourteen sections of mental models rather than tool settings: LLMs as an over-confident pair programmer, context is king, you must test what it writes, human oversight cannot be automated away.
Best Practices for Claude Code
Anthropic's official guide to working with an agentic coding tool: managing the context window, separating exploration from planning from implementation, writing CLAUDE.md files, giving the agent verifiable checks, and adding adversarial review before shipping.