Detailed Analysis
A hobbyist developer in the UK has built a feel-good news aggregator website — goodnewsforthe.uk — using a combination of Claude and Google's Gemini Flash free tier, demonstrating how consumer-accessible AI tools are increasingly enabling solo developers to build functional, production-grade applications with minimal infrastructure cost. The project runs on a LAMP stack hosted on a £3-per-month VPS, scrapes RSS feeds from UK news outlets, and uses an AI-driven scoring pipeline to filter articles by positivity and UK relevance, then rewrites clickbait headlines into plain-English summaries before publishing them. The developer, self-described as "old school," turned to Claude not only for code generation but for architectural guidance — and notably, Claude itself recommended offloading the computationally intensive LLM tasks to Gemini Flash rather than attempting to run a local model like Qwen 0.5B on the limited VPS hardware, a pragmatic and technically sound suggestion that reflects Claude's growing role as an engineering advisor rather than just a code-completion tool.
The core technical artifact shared in the post is a detailed prompt engineering document — essentially a structured editorial ruleset written in natural language — designed to guide Gemini Flash in scoring news articles on a 1-to-10 scale across two axes: UK relevance and genuine positivity. The prompt is notably sophisticated, including explicit negative examples (e.g., a BBC article about a non-UK story does not count as UK news), hierarchical decision logic (relevance check before positivity check), and structured JSON output requirements. The developer also requests geographic tagging at granular levels — county and city slugs — and category classification, all returned in a single inference call. The system is largely functional but exhibits the classic failure mode of LLM-based classifiers: occasional false positives, where articles that should score low receive inflated scores, suggesting the model is not consistently applying the rejection criteria.
The scoring reliability issue the developer raises sits at the intersection of two well-documented challenges in applied LLM workflows: prompt brittleness and inconsistent instruction-following under edge cases. Large language models, including those in the Gemini family, tend to weight fluency and surface-level positivity signals heavily, which can cause stories framed optimistically — even if they involve illness, death, or non-UK subjects — to slip past hard-coded rejection rules. Several prompt engineering strategies could address this: adding a chain-of-thought reasoning step before the JSON output (forcing the model to explicitly reason through the UK relevance check before committing to a score), introducing few-shot examples of correctly rejected articles directly in the prompt, or implementing a two-pass system where a second inference call independently audits the first score. The developer's instinct to surface this to the community reflects an increasingly common dynamic where non-ML practitioners are building LLM-powered pipelines and encountering the gap between a model's average performance and its worst-case behavior.
More broadly, the project illustrates a meaningful shift in what a single motivated developer can build in 2025-2026 without institutional resources. The combination of Claude for architectural planning and code generation, a free-tier frontier model for inference, a cheap cloud VPS, and open RSS infrastructure produces a pipeline that would have required a team and significant budget in a prior era. Claude's role here is particularly notable: rather than simply generating code on demand, it functioned as a system design collaborator — advising on model selection, flagging hardware constraints, and producing integration logic. This positions Claude less as a coding autocomplete tool and more as a junior technical co-founder, capable of scoping a problem and recommending a sensible architecture before writing a single line. The project, while modest in scale, is a concrete example of AI democratizing software development and content curation in ways that extend well beyond enterprise use cases.
Read original article →