lime The field manual · 10 patterns
Working with Coding Agents
Concrete patterns that survive real use: what to watch out for, and the habits worth building. Written for humans to adopt one step at a time; every step ends with an optional prompt that helps you apply it in your own environment.
Distilled from one practitioner's measured usage: logged sessions, the incidents that recurred, and the patterns that stopped them. v0. Step slugs are stable; chapters may be cut or reordered before 1.0.
The steps
Ordered by leverage-per-effort, but every step stands alone: start where it hurts. In a hurry, step 0 is the baseline in one paste.
- 00 The one-paste baseline
Optional baseline: durable project instructions, verification gates, and reusable review skills, set up one confirmed change at a time.
- 01 Make each correction stick
Turn corrections from one conversation into short project instructions and reusable skills that future agent sessions can follow.
- 02 Prove it before and after
Show the failure before the fix, run a cheap focused check before an expensive one, and deliver the artifact instead of a claim.
- 03 Make the agent interview you
Use a short design conversation to expose open decisions and approve the approach before the first edit.
- 04 Reviews that try to kill the work
Ask independent reviewers to find specific ways the work could fail, then verify each finding before changing the code.
- 05 Fan out, then funnel
Split large work across agents only when each part writes durable results and a final pass verifies and combines them.
- 06 Choose the right tool surface
Use a CLI for local work, an API for direct service access, and MCP when a standard agent connection adds real value.
- 07 Build lots of small tools
Turn repeated mechanical work into small deterministic tools, so the agent can spend its effort on decisions, orchestration, and glue.
- 08 Views on demand, not apps
Keep the data in structured files and generate a self-contained HTML view when a person needs to inspect it.
- 09 Turn checkable rules into checks
Keep judgment in prose, but turn mechanical rules into advisory hooks and scripts that run when they matter.
Check your work against
The guide's watch-out half lives in two numbered failure lists; every "watch out" above cites them by id: The Prose Failure List (PF-01…24, taste) · The Agent Workflow Failure List (AF-01…20, verification).
← The field manual · for agents: raw chapters at
/guides/coding-agents/*.md