LEMON AGENT

LEMON AGENT · a field guide to coding agents

You know how to chat with AI. Now learn how to work with one.

A coding agent can inspect a project, edit files, run commands, and check its work. This hands-on guide shows you how to give it a safe job, stay in control, and decide whether the result is actually right.

Start your first session See a two-minute example

Start with the course if you are moving beyond chat. Use the field manual at any skill level when you want a sharper working system.

Already working with a coding agent? Skip ahead to the field manual and the watch-out lists.

What you will learn

A better working loop, not a magic prompt.

01

Give it a bounded task

Turn “improve my site” into a job with a clear scope and finish line.

02

Stay involved without hovering

Know when the agent should inspect, plan, act, ask, or stop.

03

Make the system improve

Save corrections in project instructions, reusable skills, and small dependable tools.

Build leverage

Make every useful session improve the next one.

01

Let corrections survive the chat

Put project facts in AGENTS.md or CLAUDE.md. Package workflows as skills, with npx skills managing shared installs.

02

Prefer a clear tool surface

Use a CLI for direct local work, an API for service access, and MCP when a standard agent connection adds value.

03

Harvest small deterministic tools

Let code perform exact repeated operations while the agent handles judgment, orchestration, and glue.

Start here

Seven short lessons. One real workflow.

Each lesson gives you one idea, a decision to practice, and something you can paste into your own coding agent.

Open the 30-minute course
  1. 01
    From answers to actionsWhat changes
  2. 02
    Choose a safe first taskBound the job
  3. 03
    Write the briefGoal, scope, proof
  4. 04
    Plan before codeName the forks
  5. 05
    Watch the workKnow when to stop
  6. 06
    Verify the resultEvidence over confidence
  7. 07
    Review and saveInspect the change

Go deeper

The field manual

Patterns from measured use, written for clarity without reducing the technical depth.

10 patterns · lime

Working with Coding Agents

The core field guide: verification loops, design interviews, adversarial reviews, tool surfaces, small tools, and generated UIs. One pattern at a time, each with a prompt you can adapt.

6 patterns · tangerine

Shipping Software with Git + Coding Agents

How git and an agent compose: checkpoint commits, branches as blast radius, history the next session can read, diff-not-summary review, and recovery cheap enough to lean on.

7 patterns · grapefruit

Keeping a Vault

A personal knowledge base an agent can actually work: plain files in a repo, directories as state, an inbox habit, a daily log, and a doctor script keeping it honest.

Check your work

Use the watch-out lists as a final pass.

The Prose Failure List 24 concrete failure patterns → The Agent Workflow Failure List 20 concrete failure patterns →

Experiments: Lemon Agent also houses small tools and agent-facing services. The first, pathgrip, is a bookmark store agents can call. It runs as the operator’s own instance today; there is nothing for you to sign up for yet, and we’d rather say so than pretend. See its current status →