A design file was never the work. It was the receipt for the work, and it functioned as one because producing a good file was slow. You could not make it without having made the decisions inside it, so effort and judgment were both readable from the output.
A credible screen now costs about ninety seconds.
So the receipt stopped certifying anything, and it is still being handed in. What replaces it is a design decision record. One page, four fields, roughly ten minutes.
Field one, the decision. One sentence, present tense, specific enough to be wrong. "Suggested replies appear inline beneath the composer, maximum three, never as a modal and never auto-selected." Somebody can disagree with that, and somebody can catch a build violating it. "We improved the suggestion experience" is a summary, which nobody can disagree with and therefore nobody can check.
Field two, the constraints that produced it. Two to four rules, each written as a rule plus the observable moment it was broken. "A suggestion the user can accept in one click has an undo that is equally cheap. Broken when accepting takes one tap and reversing takes a support ticket." This is the field that makes the record work on a stranger. If the second clause is hard to write, the first clause is a principle wearing a costume.
Field three, what we gave up. The options you killed and what killing them cost. "Modal picker with six options: higher acceptance in the first session, at the cost of interrupting people who already knew what they wanted to say."
Everyone skips this field. It is the only one that proves a decision took place, because if you cannot name a real cost you had a preference, and preferences do not need records. It also pays for itself two quarters later when somebody proposes the rejected option confidently, as if it were new.
Field four, how we will know this was wrong. One signal, a rough threshold, a date to look. "People editing suggestions before sending, rather than accepting or ignoring them, means three was too few or the ranking is poor. Checked in the next round of sessions." Not a dashboard. One observable thing.
The bar the whole record has to clear: a competent stranger, or an agent, reaches the same decision from it without you in the room. If reading it leaves someone needing a conversation with you, it is a meeting summary rather than a specification.
Four triggers decide whether a decision earns a record.
Somebody will ask why more than once. You killed credible alternatives. The decision constrains what others can build later. You would be annoyed if it got quietly reversed. That last one is the most reliable detector, and mild irritation at an imagined future reversal is worth trusting.
Two filing habits decide whether any of it survives a real team. The record gets written by the person who made the call, because delegating it turns it into minutes. And it gets linked from wherever the work lives, because a record nobody can reach from the work is a diary entry. Keep them in one directory, number them sequentially, never delete. A superseded record keeps its number and points at its replacement.
One thing to do this week: take a decision you already argued for at least twice and write the record from memory. Notice which field was hard. Hard constraints mean the decision was instinct and is not transferable yet. A hard "what we gave up" means there may not have been a decision. A hard falsifier means nobody currently knows what would change your mind, including you.
The full template with the write-one triggers, filing rules, and a worked example is The Design Decision Record. The long argument is The Receipt and the Work.