Developer Builds Self-Verifying AI Rule System to Track Compliance in Real Time

·2 views

A developer discovered that none of his 15 AI assistant rules were being followed after reviewing 30 sessions, prompting him to redesign them with embedded verification markers. Inspired by Test-Driven Development, he added output tokens like [✓THINK] to each rule, making compliance grep-searchable in session transcripts. He then wrote a 350-line Python script called config-health.py that runs after every session, logs execution rates, and flags unmet rules in a pending-verifications.md file. The approach shifted rule compliance visibility from guesswork to measurable data without requiring any AI inference. The author argues the principle applies broadly — any rule, whether in code reviews or productivity systems, is merely a wish unless it carries a verifiable success criterion.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

How Data Modelling and Star Schema Improve Power BI Dashboards

Power BI is a business intelligence tool where effective reporting depends heavily on well-structured data models, not just visual elements. Data modelling involves organizing tables and defining relationships between them so that information can be combined meaningfully without duplication. Relationships link tables through shared fields, such as connecting customer names to IDs, eliminating the need to repeat data across tables. The star schema, one of the most common database structures, places a central fact table surrounded by dimension tables that provide descriptive context. A well-designed data model improves consistency, simplifies maintenance, and enables dashboards to update automatically and support informed decision-making.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How One Developer Used AI to Write 150-Word Biographies of Forgotten Women

A developer building short video biographies of overlooked notable women on a site called Tycoona turned to AI to help write 150-word scripts for 60-second reels. Using Claude's Sonnet model, early drafts repeatedly fell into common AI pitfalls such as hyperbole, generic narrative arcs, and an overemphasis on morbid details. To fix this, the developer created a dedicated 'screenwriter' persona for the AI, guided by a voice principles file emphasizing factual grounding, concrete detail, and narrative arc over lists of achievements. A structured knowledge base of validated facts — including timelines, quotes, and legacy notes — was fed into the system to keep outputs accurate. The effort required significant human editing, but ultimately produced publishable story-driven scripts suitable for social video.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Open Knowledge Format Cuts AI Hallucination Rate from 18% to 3% in Enterprise RAG

A software team behind Mattrx, a multi-tenant marketing-analytics SaaS platform, replaced conventional document chunking in their AI pipeline with a structured approach called Open Knowledge Format (OKF). Unlike standard retrieval-augmented generation, which splits documents into context-free text fragments, OKF organizes information into typed, governed knowledge units carrying metadata such as ownership, expiry dates, and explicit relationships. After rebuilding their knowledge base into roughly 11,000 OKF units, the team reported the assistant's hallucination rate dropped from 18% to 3% and stale-answer rate fell from 11% to 1.5%. Context tokens consumed per query also shrank from 14,000 to 3,500, as the system could retrieve precisely relevant units rather than broad text chunks. The approach combines vector search with graph-based retrieval, enabling the AI to answer multi-hop questions that previously went unanswerable and to enforce business rules as structured data rather than buried prose.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Green CI, Broken Page: Why Automated Tests Miss Real-World Visual Bugs

A developer discovered their landing page was visibly broken on a customer's Pixel phone — with clipped text, an unreadable button, and a missing hero image — despite all CI checks, Lighthouse scores, and Playwright tests passing. The headline was clipped due to a CSS 'white-space: nowrap' rule that prevented text from wrapping on narrow viewports, a flaw the browser could have flagged via a simple scrollWidth check. The signup button's color contrast ratio was approximately 1.4:1, far below the WCAG minimum of 4.5:1, yet no test was computing contrast on an actual rendered screen. A missing mobile image file returned a 404 only on mobile network profiles, which desktop-based devtools never triggered. The developer concluded that CI tools verify logical behavior but are structurally blind to visual rendering issues across real device viewports, and that programmatic visual checks at multiple screen sizes are necessary to close this gap.

0 comments Read more at DEV Community