Skip to content

Initialize proper agent instructions for different #17

@jdsika

Description

@jdsika

This is a minimum setup for working with agents. Never use MCP if you can avoid it. Always use skills because that uses less context. Models tend to be best up to the middle of their context window.

AGENTS.md

Pretty much every agent cli calls this differently. Most tools also allow you to put it into a subdirectory (e.g., .claude/CLAUDE.md. Ideally maintain all different types of markdowns for all the CLIs you use.
Most CLIs have some sort of \init command you can type to auto set up the top-level location of the agents file. You can also set up global files, but that is not great since you want them committed to git.

Skills

Skills are additional instructions for agents that they can find on-demand.

Use this tool to easily install skills for everything: https://openpackage.dev/docs

Rule of thumb is to not install too many skills. If you have something complex like blockchain interactions or libraries, you can possibly find a skill for that that explains how to use that.

Workflow

  1. Fresh branch, fresh context session.
  2. In plan mode, ask the CLI to plan a new feature/fix. It will usually ask questions back at least two or three times. You can also prompt it to ask about unclear things explicitly. This may depend a bit on the CLI you use.
  3. When the plan is ready, implement it by switching to build mode and just saying something like "implement it". Good CLIs (like Claude Code) usually clear the context automatically which helps the model get into the "good" context zone.

Metadata

Metadata

Labels

No labels
No labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions