How to Build an AGENTS.md File That Guides AI Coding Agents Effectively

·1 views

AGENTS.md is a single Markdown file placed at the root of a code repository that instructs AI coding agents on how to navigate and work within a project. A tutorial published on DEV Community walks developers through building one step by step, using a Python-based URL shortener API as a practical example. The file is structured into sections covering setup commands, test requirements, project layout, coding conventions, commit rules, and explicit guardrails against common mistakes. Each section is designed to give an AI agent precise, actionable information rather than vague guidance, reducing errors and broken outputs. The author argues that a well-written AGENTS.md allows an agent to understand a project's requirements in roughly thirty seconds, making it one of the most impactful files a developer can add to a repo.

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)

Developer Builds Open-Source Tool to Preserve AI Coding Context Across Sessions

A developer noticed that while Git tracked code changes, the reasoning behind those changes — failed attempts, agent warnings, and deferred tasks — was lost across disconnected AI sessions. Switching between tools like Codex, Claude Code, and ChatGPT meant each new agent started from scratch, forcing repeated re-explanation of prior decisions. To solve this, the developer built ContextVault, a local-first, open-source memory layer that captures and stores context from browser LLM chats, terminals, and coding-agent sessions. The tool includes a Chrome extension, an npm CLI, and a desktop app, all storing data locally in Markdown without requiring any account or external backend. ContextVault lets developers or agents query past decisions, problems, and tasks, and generate focused context packages to hand off seamlessly to the next AI session.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Unitree's $4,900 R1 AIR Humanoid Robot Goes on Sale Globally via AliExpress

Chinese robotics firm Unitree began shipping its R1 AIR humanoid robot globally in April, listing it on AliExpress for $4,900 — making it one of the most affordable bipedal robots ever sold to the public. The 4-foot-tall, 55-pound machine offers 20 degrees of freedom and can run, perform flips, and execute martial arts moves, with an upgraded $5,900 model adding greater articulation and depth perception. Unitree shipped approximately 5,500 units in 2025, vastly outpacing Western competitors like Figure AI and Tesla, which each shipped around 150 units combined. However, the base model lacks functional hands, has roughly one hour of battery life per charge, and many of its widely circulated demo videos have been noted as scripted or remotely controlled. Analysts describe the R1 as an accessible experimental platform rather than a capable household assistant, drawing comparisons to the Raspberry Pi's role in democratising computing.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds Telegram bot that estimates calories from food photos

A software developer created a Telegram bot called SnapPlateBot that identifies food from photos and returns calorie, protein, fat, and carb estimates. The bot, built with Python and aiogram, uses a vision language model to analyze dishes and estimate portion sizes without storing any images. Early versions misidentified foods — notably labeling borscht as berry compote — until the developer improved accuracy by feeding the model physical context rules rather than longer dish lists. Users can interact via photo, text, or voice, and payments are handled through Telegram Stars, requiring no signup or app store. The bot is currently free for limited use and the developer is seeking feedback on non-Russian cuisines to improve its performance.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Narrative-Driven Reviews Outperform Traditional Software Dev Metrics

Traditional software development metrics such as commit counts, velocity charts, and bundle size are widely used but often fail to capture the full context of a developer's work. These lagging indicators reflect past actions without revealing the decisions, reversals, or thought processes behind them, leading developers to miss critical patterns. A narrative-driven approach — reviewing commits, pull requests, and issues as a sequence — surfaces the story behind the numbers and exposes inefficiencies invisible on standard dashboards. One developer found through narrating their own weekly work that repeated decision reversals, a pattern hidden in metrics, were significantly affecting productivity. The article argues that for complex, context-dependent work, structured narrative reviews offer more actionable insight than conventional dashboard metrics.

0 comments Read more at DEV Community