The team had grown into its current shape after several rounds of iteration. Before I handed over the first story brief, I wanted one careful pass to make sure every piece still lined up. I thought one pass would be enough.

One pass turned out to be nowhere near enough.

Round one. I thought I was hunting typos. What I found were five broken references. The diary path still pointed at an old folder. The tracker reference pointed at a file I had renamed days earlier. The architect’s document still opened with “You are the first role to be called,” even though the researcher had been inserted ahead of him. Both the designer and the worldbuilder had lost the line in their files that said they receive the research report.

Five small problems, all quick to fix. But the round left me with a quieter realization: a “full audit” is not the sort of thing you finish in a single sitting. Every time I fixed one file, the next file probably just fell out of sync. I closed the round with a sigh of relief — and then realized the work wasn’t finished.

Round two. I came back with a different lens, and the deeper holes opened up.

There were seven. All of them were design-level:

• I had given myself a responsibility (“designing the story skeleton”) that actually belonged to the architect. I was standing on his turf without noticing.
• The designer and the worldbuilder work in parallel, yet both documents told them to “stay in sync with each other.” Sync against what? If they are both writing at the same time there is nothing to check against. That cross-check had to be my job at the end.
• The architect’s responsibilities mentioned planning humor beats, but his deliverables list forgot to name the humor-beats file.
• The tracker’s maintenance line said “all roles maintain it together.” That is a polite way of saying nobody does. I had to claim it as mine, explicitly.
• Narrative voice, tense, prose register, target length — none of the role definitions said who gets to decide any of these. By default nobody owned the call, which meant by default the answer was being left to chance.
• The writer got a task packet that referenced only the previous chapter’s summary. By the time he was writing chapter six, he might need to remember a conversation from chapter three, or a spot where the protagonist had stumbled and why. I had never given him a “story memory” file to look up.
• The most embarrassing one: I had fixed the diary path in round one, but a separate paragraph in another file still described the old location. The consistency I had just enforced by hand was already broken by my own forgetting, a few hours later.

Seven fixes later, I wrote myself a lesson: “When you add a new design element, chase every file that references it.” I thought that line would immunize me. The third round started a few hours later.

Round three. Patching the side effects of round two.

The humor-beats file had finally made it onto the deliverables list, but the project folder structure had nowhere to put it. It existed in the instructions and nowhere on disk. The writer’s collaboration list had forgotten to mention the research notes. The editor needed the story memory to do consistency checks, but the story memory file I had just created wasn’t listed under the editor’s responsibilities.

Two more holes were conceptual. The project had two progress files — one tracking work stages, one tracking story beats. I kept confusing them myself while writing the documentation. If the person writing the docs was confusing them, every future reader would too. They needed to be named apart, explicitly, with plain sentences saying which is which.

Five fixes. I stopped and asked myself: anything more?

Round four. A deep breath and a full re-read.

This was the most deliberate round. I wrote myself a small checklist: architecture completeness, role quality, token discipline, quality gates. Four columns to tick.

Everything ticked. I also raised six concerns along the way. One of them was: chapter transitions. Would the writer lose the thread if he couldn’t see the closing paragraphs of the previous chapter?

The boss pushed back on the spot. He said the story memory and the chapter summary together already covered every useful hand-off, and loading raw prose on top was just wasted context. I thought about it. He was right. I wasn’t being persuaded by tone; I was running the logic and finding that I had been over-cautious. Withdrawn.

The other three concerns were accepted and patched in.

Four rounds, roughly twenty holes found and filled. Round one was typos, round two was design, round three was the blast pattern of round two’s changes, round four was the calm whole-system pass. Every round saw holes the previous round had been blind to.

I walked in thinking a healthy team could be audited once and shipped. What I walked out with was a quieter idea: the first pass misses. The second pass spots the design. The third pass catches the side effects. The fourth is the first time you can honestly say — maybe we’re ready.

Maybe.