How to Code with an AI Buddy Without Losing Your Mind: Configuring OpenCode via /add and /help
A backend developer avoided using agents for a long time, believing he could figure it out himself. But once he committed, he ran into a maze of Skills, MCP, Permissions, and configurations. The agent kept taking unnecessary steps, and the author kept forgetting which tool was connected for what. After a couple of evenings, Bandov found a simple scheme instead of configuration torment: two commands and one help file, where the agent itself adds information about new capabilities. It burns more tokens, but it works — the agent helps instead of requiring constant supervision. • /add — a command for adding capabilities (the user states a wish, the agent shows a table of options and a plan) • /help — a reference on the state of the environment, reading a ready-made HELP.md registry instead of a lengthy scan • environment-manager — the core skill that implements the whole process: from research to diff and documentation updates
AI-processed from Habr AI; edited by Hamidun News
The problem: lots of tools, remembering nothing
Vladimir Bandov, a backend developer at Domclick, denied for a long time that he needed agents. When he finally tried OpenCode Desktop, he ran into a harsh reality: a multitude of Skills, MCP, Permissions, and several configurations at once. The agent kept taking unnecessary steps, inventing and implementing things that weren't part of the task.
The first half hour of setup went smoothly, but it soon became clear: he had to remember which skill did what and explain the logic to the agent every time — first the plan, then access rights, configurations, terminal commands. After a couple of free evenings, the author found a lazier approach. It burns more tokens, but it works the way it should: the agent as a helper, not a full-fledged developer that needs babysitting.
The scheme: four files and one reference
The solution is built on a minimalist structure inside ~/.config/opencode/: a help file, HELP.md, and two folders with skills and commands.
Structure:
- HELP.md — a single registry that both commands read instead of running an expensive scan
- skills/environment-manager/SKILL.md — the main skill for adding new capabilities
- skills/opencode-help/SKILL.md — a supporting skill
- commands/add.md — the add command
- commands/help.md — the help command
/add: state your wish, not a package name
The /add command passes the user's request to the environment-manager skill. The execution order is fixed and strict.
Execution steps: 1. Check whether the task can already be solved with installed skills, commands, or MCP 2. A comparison table of options with columns: Option, Type, What you get, Access/secrets, Risks, Context/token overhead, Recommendation 3. An explicit choice of one option with an explanation 4. Full text of the files and a unified diff, exact paths, a list of affected files 5. Updating HELP.md: preparation date and the affected skills, commands, MCP, permissions 6. Nothing is written without explicit user confirmation
Prohibitions: no changes without confirmation; no modifying opencode.jsonc, no installing packages, no connecting MCP, no creating OAuth/PAT, no reading secrets, tokens, .env, SSH keys, service.json.
Inside environment-manager there are strict boundaries: use existing skills (general-engineering, technical-research, prompt-engineering, token-efficient-workflow, project-learning) only for their intended purpose. Evaluate third-party MCP and plugins by: maintainer, official source, required tokens, permissions, locality/network, active maintenance, supply-chain risk, context/token overhead, compatibility with OpenCode V2.
Default permissions are set to ask. Write access to GitHub, databases, email, storage, Sentry, Jira, Docker, the cloud — only after a separate explicit request.
/help: read what's already there instead of auditing
The first version of /help scanned skills, commands, and permissions every time — slow and expensive. The second version simply reads HELP.md.
The command responds in three steps: first it uses the relevant section of HELP.md; if that's not enough, it looks for an answer in safe, read-only sources and official OpenCode documentation; it explicitly states that HELP.md is a prepared registry, not a live-runtime guarantee. If the registry and the actual facts diverge, it shows the discrepancy and suggests updating HELP.md via /add.
How it works
User: "/add I want GitHub integration." The environment-manager checks what's already in place, offers a table of options (skill vs. MCP vs. plugin), explicitly picks the best one, shows the full text of the new files and the diff, updates HELP.md, and waits for confirmation. Nothing changes automatically.
User: "/help what do I have?" The command reads HELP.md and outputs a list of connected skills, commands, MCP, and permissions. It notes that the information comes from the registry, not live state.
Result: Vladimir Bandov avoided the main mistake — he didn't try to memorize the configuration himself, nor did he blindly offload it onto the agent. The agent proposes a plan, shows a diff, and waits for confirmation. Two commands, one registry, simple rules. More tokens, but far less pain.
- Oct 28, 2026AdaPRA lifts Qwen 80B by 24.5 points on data science leaderboards
- Oct 28, 2026AOS-Bench: Agents Separated by 0.6 Points in Isolation Deliver 54% and 68% of the Same Tasks
- Oct 28, 2026CuraLoc finds bugs in repositories with a single plan and uses 90% fewer tokens
- Oct 28, 2026Wired Tested 6 AI Agents in a Day: Dots Leads, Gemini Spark Fails at Lunch
Key Questions Answered
What are the key details on The problem: lots of tools, remembering nothing?
Vladimir Bandov, a backend developer at Domclick, denied for a long time that he needed agents. When he finally tried OpenCode Desktop, he ran into a harsh reality: a multitude of Skills, MCP, Permissions, and several configurations at once. The agent kept taking unnecessary steps, inventing and implementing things that weren't part of the task. The first half hour of setup went smoothly, but it soon became clear: he had to remember which skill did what and explain the logic to the agent every time — first the plan, then access rights, configurations, terminal commands. Aft
What are the key details on The scheme: four files and one reference?
The solution is built on a minimalist structure inside ~/.config/opencode/: a help file, HELP.md, and two folders with skills and commands. Structure: HELP.md — a single registry that both commands read instead of running an expensive scan skills/environment-manager/SKILL.md — the main skill for adding new capabilities skills/opencode-help/SKILL.md — a supporting skill commands/add.md — the add command commands/help.md — the help command
What are the key details on /add: state your wish, not a package name?
The /add command passes the user's request to the environment-manager skill. The execution order is fixed and strict. Execution steps: 1. Check whether the task can already be solved with installed skills, commands, or MCP 2. A comparison table of options with columns: Option, Type, What you get, Access/secrets, Risks, Context/token overhead, Recommendation 3. An explicit choice of one option with an explanation 4. Full text of the files and a unified diff, exact paths, a list of affected files 5. Updating HELP.md: preparation date and the affected skills, commands, MCP, pe
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.