Articles from the Paqad project on running AI coding agents within project rules: what loads, what a check can enforce and what still needs review.
How to get an AI assistant to investigate a fault instead of narrating a guess: sort reproduced facts from reported and assumed ones, test competing explanations with experiments that could rule them out, and agree what proves a root cause,…
Which design choices need a written record, where to keep it so people and coding agents find it while they work, how to supersede it without losing the reasoning, and how to tell an approved choice from an agent's plausible…
How to onboard an engineer into a team that works through AI agents: what to ask the repository and what to ask a teammate, a guided first change the newcomer owns, and an agent-off task that shows understanding rather than…
How to keep guides and examples true when AI-assisted changes keep the tests current: list what each change made untrue, run documented examples against the code, stamp each guide with the commit it was checked against, and a demonstration of…
Which parts of a database migration to hand an AI coding agent, which to keep, how to give it a database it is allowed to lose, and a rehearsal in which a migration passed every usual check while deleting 198…
Which Flutter work to hand to an AI agent, what to decide first, the widget tests that catch faults the analyzer missed in a small demonstration, and how to time the whole task, debugging included.
One older benchmark found models did better in React than Vue, and documentation in context moves agent results a lot. Diagnose failures, run a representative ticket in your Vue app with better context and checks, then weigh what still fails…
Read the installed Next.js version and router from the repository, give the agent the documentation bundled with that version, back it with a build, lint and a search for retired APIs, and change the rules in the same pull request…
Give the agent the application's own account of its routes, models, policies and conventions, brief the feature around the existing permission and data rules, write the forbidden-user test first and review the diff where Laravel hides behaviour.
How to attribute coding agent spend to tasks and attempts, tell useful retries from wasted ones, match the model to each step without weakening verification, and judge the result by cost per accepted change.
