agents-best-practices
A provider-neutral Agent Skill that helps design, generate an MVP blueprint for, audit, refactor and explain agentic coding and workflow harnesses, covering tool permissions, planning, context and memory, evals and production readiness.
agents-best-practices is an Agent Skill, usable with Codex, Claude Code and other compatible agents, for working on the design of an agentic harness itself rather than on a specific coding task. It activates when a conversation touches agent architecture, tool permissions, environment-adaptive tools, speculative tool execution, planning mode, workflow orchestration, context and memory, skills, connectors, troubleshooting, observability, evals, prompt caching or production readiness, and the same core practices are meant to apply beyond coding agents to research, support, operations, sales, finance and other workflow-automation agents.
It supports several concrete tasks: generating the smallest useful, production-safe MVP agent blueprint for a given domain; auditing an existing agent that has become brittle, expensive, too broad or hard to debug; designing tools, permissions and connectors so an agent can use real systems without turning every API into an unsafe broad tool; and adapting an agent to a late-bound tool environment that varies by customer or deployment. Each case points to a specific reference document in the skill's own material.
On each new task, the installed skill checks the repository's latest revision and updates itself to a consistent snapshot when it is stale, while preserving local changes.