DESIGN.md: the file after AGENTS.md
Matt Connor · 9 October 2026
Read the originalMatt Connor, chief executive of the hosting company SSD Nodes, proposes a DESIGN.md file that records settled decisions so that coding agents stop undoing them.
Matt Connor founded and runs SSD Nodes, a company that rents virtual private servers. He describes coding agents that replace patterns they do not recognise, such as a hand-written cache, while the tests still pass. His answer is a short file at the root of the code repository that records what must stay true and why.
- The file lists the rules that must stay true, each with its reason, and the rejected alternatives with conditions for revisiting them. It also lists high-risk boundaries, such as schemas and deploy settings, and a vocabulary mapping code terms to business terms.
- AGENTS.md tells an AI agent how to work in a repository, while DESIGN.md tells it why the code is shaped as it is.
- Connor advises updating the file in the same commit that changes a decision, and adding a line only when an agent makes a mistake that one sentence would have prevented.
- He advises testing the file by giving the agent a task that touches a protected rule and checking whether the agent cites it.
A team can make the person who changes a settled decision responsible for recording it in the same commit. Connor names no owner for the file, so who approves changes to it remains open; that observation is an inference.
List the rules a coding agent must not break in one file in the code repository, then give the agent a task that touches a rule and check whether it cites that rule.
Keep one file in the code repository that lists the team's settled engineering decisions for coding agents, and update it in the same commit that changes a decision.
Derived by Working Surface from the article; more in the Playbook. Source line: DESIGN.md tells it why the code is shaped that way, so it stops undoing your decisions.
11 October 2026: AI agents depend on written context that goes stale, and the teams in these pieces review agent changes more firmly than they keep that context current.