Anbeeld/AGENTS.md

README preview

# Anbeeld's Global AGENTS.md / CLAUDE.md

A global instruction file for coding agents that fixes what they most commonly get wrong: jumping to code without evidence, expanding scope unchecked, or shipping changes that haven't been validated.

Every instruction came from the same pipeline: observe what agents get wrong, compare it with public examples and [recommendations from industry leaders](https://github.com/Anbeeld/AGENTS.md/commit/f796d26c30402dc0aab12cb02e9e448504047d98), test my own solution when needed, then add or tighten a rule.

Battle-tested in daily use, tuned across multiple agents and models. Use the full set, borrow what you need, or build something different from it.

[![Support my work!](https://anbeeld.com/images/support.jpg)](https://anbeeld.com/support)

## What it does

- No wrong-path work: asks before guessing on ambiguous tasks, investigates proportionally to risk, and doesn't improvise when goals conflict.
- Hard boundaries: no fabrication, no gamed verification, no exposed secrets, no destructive commands without explicit confirmation.
- Structured execution order: explore, gather evidence, scan skills, implement, verify. Independent work runs in parallel when it helps, either through parallel tool calls or a batch of two or more subagents.
- Keeps changes small. No scope creep, no unprompted refactors, no new dependencies without checking what's already available.
- Protects existing code: tests match the risk level, behavior isn't silently changed, errors aren't swallowed. If verification fails, the agent makes one targeted fix when the cause is clear; otherwise it stops and reports.
- Validates before finishing: the agent confirms the change solves the problem, validation ran, no side effects or secrets slipped in, and the format is correct.
- Ops work in a fixed order: environment, services, configs, logs, then fix. Config validated before reload, service names from project instructions, not guesses.
- Short answers by default. No filler, no sycophancy, no restated requirements. For reviews and debugging: findings with references, then conclusion, then approach.

View full repository on GitHub →