Перейти к содержимому

Agent Skills: Procedures Your Coding Agent Loads Only When Needed

BigIron AI

0:00 / 0:00

Agent Skills: Procedures Your Coding Agent Loads Only When Needed

4 просмотра · 1 день назад
BigIron AI
15 подписчиков
4 просмотра · 1 день назад
An instruction file like AGENTS.md or CLAUDE.md is loaded into every session, so it should hold short facts. Twenty-step procedures that matter once a month cost context every day if they sit there. Skills are the answer most coding agents have converged on: a folder with a SKILL.md file, holding a short description the agent always sees and a body it loads only when the task matches. The format is the open Agent Skills standard, supported by Claude Code, Codex, opencode and a growing list of other tools. This video follows the bigiron.cc guide through the format, where each agent looks, descriptions that trigger, side-effecting skills you keep under your control, and keeping one copy in sync and validated. What's covered: • Facts in the instruction file, procedures in skills • When you don't need skills at all • Progressive disclosure: about 100 tokens per skill until a task needs it • The six portable frontmatter fields, and why extra fields break strict tools • Where Claude Code, Codex and opencode look for skills (as of late September 2026) • Why Claude Code doesn't read .agents/skills • opencode loading Claude Code's skills: about 750 tokens per session • Listing budgets: 2% or 8,000 characters in Codex, 1% in Claude Code • Descriptions that trigger: what, then when, in the words you type • A real 1,013-character description, 11 inside the 1,024 limit • Testing triggers in Claude Code and Codex • Writing the body as a runbook, with scripts for anything checkable • Manual-only skills, and why that is not a permission system • allowed-tools grants, it does not restrict; read third-party skills first • Symlinks, a sync script with a drift check, or a plugin • Validating skill folders, including stray backup files ⏱️ Chapters 0:00 Agent Skills 1:33 How Skills Load 2:47 Where Each Agent Looks 4:18 Descriptions That Trigger 6:03 Runbooks and Side Effects 8:17 One Copy, Kept Honest 📖 The full guide (the complete add-host skill, check-inventory.sh, the sync-skills.sh sync tool, the check_skills.py validator and a troubleshooting table): → Agent Skills for a Homelab: SKILL.md Procedures Agents Load on Demand: https://www.bigiron.cc/guides/agent-s... 🛠️ Gear (affiliate — same price to you, supports the channel): • Mini PC for an always-on homelab agent box → https://www.amazon.com/s?k=mini+pc+32... • NVMe SSD for a dev workstation → https://www.amazon.com/s?k=nvme+ssd+2... • Mechanical keyboard for long terminal sessions → https://www.amazon.com/s?k=mechanical... As an Amazon Associate we earn from qualifying purchases. Affiliate links — same price to you, supports the channel. 👉 Subscribe for self-hosted AI + homelab, with new videos every day. #AgentSkills #ClaudeCode #Codex #opencode #AIAgents #CodingAgents #Homelab #Automation #SelfHosted #AI