63 posts tagged with "Ai"

LLM prompts are full of repeated English grammar. Every "If a module exists, then recommend it. Do NOT reinvent the wheel." burns tokens on words the model already understands. I built CPF (Compact Prompt Format) to replace that grammar with operators and abbreviations that LLMs decode natively — cutting token costs by 30-50% with zero runtime dependencies.

Dries Buytaert, the founder of Drupal, recently addressed a growing problem: AI-generated contributions are flooding open-source projects with low-value reports and patches that lack expertise. The human bottleneck — the reviewer — becomes the point of failure for the entire ecosystem.

Drupal 12 lands mid-2026. The deprecation list is long, the database API is changing, and most teams will not discover their exposure until the upgrade breaks something in production. Static analysis catches the problems early. But terminal output is not enough when you need to share findings with project managers, clients, or stakeholders who do not read diffs.

The Drupal 12 Readiness CLI now ships with an HTML report generator, GitHub Actions CI across three PHP versions, and coverage for all 32 deprecated database API functions.

AI is making vulnerability discovery faster and cheaper. That is the easy part. The hard part is what happens next: an open-source maintainer with limited hours receives a flood of security reports and must decide which ones deserve immediate attention, which are false positives, and which can wait.

I had two collections of plain arrays and needed to associate them relationally, but I was not working with Eloquent models. Laravel's Collection class is powerful, but it has no built-in way to express a one-to-many relationship between two arbitrary datasets. I built laravel-collection-has-many to fix that.

Gemini Ollama CLI Bridge chains a local Ollama model with Google's Gemini into a two-stage code analysis pipeline. The first version piped data to Gemini through a shell subprocess -- fragile, hard to test, and impossible to cache. This upgrade replaces the shell integration with the google-generativeai Python SDK, adds result caching, and grows the test suite from 4 tests to 22.

The first version of Drupal AI Content Impact Analyzer was a minimal proof-of-concept: a single scoring function, no tests, no documentation worth mentioning. It proved the idea worked. It did not prove the idea was useful to editors who need to know exactly why a piece of content scores the way it does. The upgrade rewrites the core into a 6-dimension scoring engine that tells editors precisely what to fix.

The honeymoon phase of "generate everything with AI" is officially over, as major platforms like WordPress and Cloudflare are now forced to build guardrails against the resulting tide of low-quality "slop."

The AI model landscape just shifted again with the simultaneous drop of Opus 4.6 and Codex 5.3, and for once, the "System Card" is more interesting than the marketing splash page.

The Drupal AI Hackathon: Play to Impact 2026, held in Brussels on January 27-28, was a pivotal moment for the Drupal AI Initiative. The focus was on practical, AI-driven solutions that enhance teamwork while upholding trust, governance, and human oversight.

The most compelling challenge: creating AI Agents for Content Creators — not simple content generation, but agentic workflows where AI acts as collaborator, researcher, or reviewer.

As AI agents become part of the Drupal workflow, you need infrastructure that makes automation traceable, predictable, and composable. I built four small Drupal modules that together form an MCP Toolkit for agent-driven site management.

The first version of drupal-entity-reference-integrity could scan a Drupal site and report broken entity references. That was the easy part. The hard part -- actually fixing them at scale without taking the site down -- was missing. This upgrade adds auto-fix mode, batch processing, and the test coverage needed to trust both in production.

I built this proof-of-concept to explore how a codex-style agent can assist with routine Drupal maintenance tasks -- dependency updates, configuration checks, and basic cleanup. The kind of work that is constant, easy to overlook, and expensive to do manually at scale.

The original POC had three actions -- create a node, read a node, list nodes. Enough to prove the concept, not enough to be useful. A real content operations agent needs to modify, delete, query users, browse taxonomies, and handle media. It also needs to survive flaky networks without crashing.

I built drupal-content-audit as a lightweight way to inspect and report on content in a Drupal site. It focuses on surfacing what content exists and how it is distributed, giving a quick snapshot that is easy to share with stakeholders.

I built a DDEV add-on that runs Playwright checks against local MCP endpoints using internal DDEV DNS names like web. This solves flaky localhost-based testing and gives one command for repeatable MCP smoke tests.

I built drupal-ai-module-generator-deepseek-mcp to take the repetitive, error-prone parts of module setup -- info files, boilerplate, and consistent structure -- and make them fast and repeatable. It fits naturally into agent-driven workflows where you want consistent Drupal modules without losing time to manual setup.

I built drupal-ai-views-agent because Views is powerful but repetitive. Filters, relationships, display modes, and fields add up quickly. Describing what you want in plain language and getting a valid Views config back saves the clicking and lets teams focus on validating data requirements instead of wrestling with UI details.

The first version of drupal-claude-agent-sdk-runtime gave you session management and result objects, but tools were mocked. You could simulate agent workflows inside Drupal, but you could not actually execute anything. That ceiling had to go.

If you install "skills" from a public marketplace, you are installing trust. I built a static scanner that scores a skill bundle before it touches my machine.