Specs, plans and checkpoints
Those conditions let us reject an implementation that satisfies only half the request.
Links containing ?t= open the video at a specific second.
The ideas to retain
A request is not yet a contract
Human requests usually express intent rather than every invariant an autonomous run needs.
Persistent instructions and a task spec solve different problems
Agentic repositories can carry persistent guidance in AGENTS.md, CLAUDE.md, GEMINI.md, or equivalent mechanisms.
A plan is a hypothesis, not the authority for the task
Current coding agents expose planning as an explicit operating mode. The official Codex Plan Mode template allows non-mutating actions to gather evidence and asks for a decision-complete…
Jump directly to a section
Read the reviewed transcript
This video has no narration. This transcript reproduces its visible explanatory text, not an invented voice.
From a request to a test
“Add JSON” leaves the format and the previous behaviour undefined.
The contract specifies a valid object with --json and unchanged text without it.
Those conditions let us reject an implementation that satisfies only half the request.
A stable rule, a specific goal
Repository instructions say how to work: do not change the public API without permission.
The spec defines this task: add --json without breaking the existing output.
Checking off a plan item changes neither obligation.
Dependencies determine the order
Agree on the schema first. The parser and serializer can then proceed.
The parser is ready, but the serializer is missing. The end-to-end test remains blocked.
Only when both pieces are ready can the complete test run. Finishing one substep is not enough.
The plan can be wrong
The plan assumes changing the serializer is enough.
The test shows that the parser rejects --json before it reaches the serializer.
Revise the plan to address the cause. A working hypothesis is not evidence of success.
Save evidence, not just notes
The checkpoint stores revision A, completed tests and pending decisions.
A parser edit creates B. Dependent tests now need to run again.
On resume, reconcile reality with the checkpoint. A nearby PASS does not carry over automatically.
The task requires the whole contract
Valid JSON is one condition, not the whole task.
The previous output and the permitted scope must also be preserved.
Delivery requires all mandatory conditions together. One green substep is not enough.
