atlas
harness
practices
field guide
rev. 4
Every practice here has been installed on a real codebase and handed to the client’s own team to keep running. Four layers decide whether a coding agent is useful or dangerous — read them in any order, or follow a playbook.
01 knowledge9
what it reads before it writes anything
without it, every session starts by guessing your conventions
02 capabilities9
the workflows it can run on demand
without them, each person re-prompts the same job from scratch
03 connectivity1
the systems it is wired into, and on whose credentials
without it, an agent can describe the fix but never apply it
04 governance7
the limits it cannot cross without a human
without them, speed nobody is willing to actually use
index
Click a practice to read it.
Chapter 0 scores your own harness in two minutes.
1knowledge
1.1Product vision1.2Product requirements definition (PRD)1.3Tech approach1.4Tech audit1.5RFC1.6Inception1.7How to write acceptance criteria that developers love1.8Software development as a system1.9Decoding direct sales: a software engineer's perspective2capabilities
2.1Automated acceptance testing strategy2.2Use Cases, checklists and automated test cases2.3Reporting a test result2.4Feature flags2.5Feature flags in a continuous delivery pipeline2.6Decouple code deployment from feature release2.7Work in small batches2.8Git branching strategy2.9Design systems - Why, what and how3connectivity
3.1Monitoring a real-world production app