alessiomarcone/agent-fieldbook

{"i": 0, "repo": "alessiomarcone/agent-fieldbook", "stars": 3, "lang": "Python", "ok": true, "char_count": 1750, "text": "Agent Fieldbook is a community project that takes official training material from Anthropic and OpenAI and converts it into practical, source-backed instructions that AI coding agents like Claude and Codex can use during real work. The project addresses a specific gap: you can use an AI agent every day, then take an official course and discover that the structure, terminology, or mechanism the agent used differs from what the course recommends. The fieldbook puts those operating lessons into the workflow itself, so you do not need to complete every official course before your agent benefits from its guidance.\n\nThe project packages this guidance as 10 focused skills with names like navigate, prompt, verify, and subagent. Each skill has a specific contract. The navigate skill helps choose one smallest adequate mechanism for a task. The prompt skill turns a rough objective into a bounded, verifiable task prompt. The verify skill labels claims by evidence and uncertainty. The project includes direct links to 21 Anthropic courses, 30 Anthropic tutorials, and 12 verified OpenAI learning resources, including three formal Academy courses.\n\nThe project also includes 12 distilled knowledge cards for Claude. These cards cover procedures, failure modes, and checks. The repository runs zero runtime credentials, hooks, background processes, or external servers. It is not a transcript archive or a replacement for the courses. It is an execution companion: concise instructions for applying stable lessons, direct links for deeper study, and explicit verification whenever product behavior may have changed.\n\nInstallation differs slightly between Claude and Codex but follows the same pattern: add the plugin marketplace, then install the knowledge-pack plugin. You can also copy a single standalone skill into your local skills directory. Claude and Codex share operating principles but not configuration formats, so every skill selects one product branch before using commands, manifests, permissions, or agent fields. Shared concepts are normalized and unsupported fields are never translated by analogy.\n\nThe full README is longer than what was shown.", "error": null}