FIELD NOTE / INSTAGRAM
The green UI has five blank cells.
The short film, the complete written thought, and the evidence behind it.
The Instagram edition will be linked here after its public post is verified.
The green UI has five blank cells.
Video caption
The green UI has five blank cells.
The team re-prompts, retests and redeploys to find the rest.
One explicit question and one observable case per cell.
Do not let a model-written checklist grade itself.
#EricFieldNotes
Full written post / accessible read
A screen that looks finished invites a fast signoff. But it does not tell you which record is authoritative, who may change it, how a retry behaves, how old data migrates, or how operations recover a half-finished action.
In an illustrative approval workflow, the first task succeeds. Then an old record enters after a migration and the new code treats it as approved. The expensive work is not necessarily hand-recoding; it is reconstructing the state model and hunting for neighboring holes.
Draw five rows before implementation. For data, identify the authoritative record. For retries, replay after a lost acknowledgment. For permission, attempt a forbidden transition. For migration, load an old version. For recovery, interrupt the worker mid-action and reconcile.
Keep those fixtures and expected outcomes in a protected acceptance harness. Let agents implement the UI and tests, but demand the independent five-row readback before signoff. Do this because a beautiful screen cannot reveal the transitions it never traversed.
#EricFieldNotes
Evidence and boundary
On-screen label: ILLUSTRATIVE REVIEW MAP. Illustrative cases are not measured incidents. Research papers and vendor documents support the stated mechanism only within their studied or documented scope.
- SWE-bench Live paper (S22)
- SWE-bench Pro Verified paper (S23)