Earlier this year we were given $2,000 in Anthropic credit, and we treated it like fuel in a tank. Every session started with a calculation. Is this task worth the tokens? Should this run on the cheap model? Can we get away with a smaller context window?
That arithmetic shaped everything about how we worked. It decided which problems we pointed AI at, which we solved by hand, and how much of a codebase we were willing to let a model read. Scarcity was the organising principle, and like most organising principles, we stopped noticing it was there.
Then the ceiling came off. We now run unlimited Fable Max, and the calculation that governed our daily work simply disappeared.
What happened next is the part worth writing about, because it was not what we expected.
The bottleneck moved
We assumed that removing the capacity constraint would make us proportionally faster. More tokens, more parallel work, more output. That is not what happened.
Within about a week it became obvious that tokens had never been the real limit. They had been the visible limit, which is a different thing. Once they stopped binding, the constraint moved somewhere less convenient: us.
Specifically, it moved to the handoffs. A model with a large context window and no budget pressure will happily work through an enormous problem in one sitting. But that sitting ends. The session times out, the context fills, the person driving it goes to bed. And everything that model understood about the problem — the decisions it made, the dead ends it ruled out, the constraints it discovered halfway through — evaporates.
When tokens were scarce, sessions were short and we carried the context in our heads. When sessions ran long and deep, we could no longer do that. We were losing more to forgotten context between sessions than we had ever lost to rate limits.
So we rebuilt the workflow around the new constraint rather than the old one.
Four things we changed
Exit interviews. Before ending any substantial session, we now prompt the model to summarise what it did, what it decided, and what it would tell the next instance picking this up. Not a commit message — an actual handover briefing. Malaz runs this on every Fable session that touches significant logic, and the summary becomes the opening context for the next session.
This sounds trivial. It is the single highest-leverage change we made. The alternative is a new session starting cold on a problem that was three-quarters solved, rediscovering the same dead ends at full speed.
Session stability in the browser. Long-running work kept dying to background timeouts when the driver switched tabs. Moving active sessions into the Claude Chrome extension fixed it. The work survives you going to look at something else, which matters enormously when a single build might run for hours.
Small, unglamorous, and it recovered more hours per week than any prompting improvement we have made.
A hard line between logic and interface. This one came out of repeated frustration. The models are genuinely strong at architecture — data schemas, backend functions, complex conditional logic, the sorting and state problems that eat a developer’s afternoon. They are, in our experience, consistently poor at user interface. Layouts come back cluttered. Hierarchy is wrong. The design is technically functional and visually incoherent.
Rather than keep fighting it, we made it a rule. We prompt for logic and architecture only, and explicitly restrict the model from generating frontend components. The structural foundations go to Wasim, who builds the interface natively. The AI does the part it is excellent at and is not asked to do the part it is not.
This is the change most teams resist, because it feels like giving something up. It is the opposite. Constraining the model to its actual strengths made the output dramatically more usable.
Two-stage handoff between people. We split builds across humans at different phases. I take the first pass to generate initial structure and unblock the project — often working directly from a client’s scope document, in one case a 125-page specification fed through in a single stretch. Malaz then takes over for deep technical execution.
The first pass is fast and rough and its job is to remove ambiguity. The second is careful and its job is to make the thing real. Separating them meant neither phase waits on the other.
What we deliberately do not do
We do not let the AI run unsupervised on anything that reaches a client. Wasim monitors output on active builds. Hiba runs a full QA pass on every release, logging each break directly into the task board, and nothing ships to a client without that pass.
That is not caution for its own sake. We published a detailed account recently of an agent of ours that reported five consecutive runs of improving results while measuring nothing at all. The numbers were real. The link between the numbers and reality had quietly broken. An agentic system will always find something to report, and the report will always sound like progress.
So the discipline scales with the capacity, not against it. More AI throughput means more verification, not less. We removed the token ceiling and increased the QA burden at the same time, on purpose.
What this means if you are hiring us
The practical outcome is compression. Work that used to take a sprint now takes days, not because we type faster but because the expensive parts — architecture, complex logic, scope ingestion, the first structural pass — collapse from days into hours.
What does not compress is judgment. Deciding what to build, spotting when an output is subtly wrong, knowing which client request is scope creep and which is the actual requirement surfacing late. That work stayed exactly where it was, and it is now the majority of what our people do.
This is the broader pattern we keep arguing for: AI adoption is not about reducing headcount but about raising what a small team can execute. A team of four that has restructured around these tools can genuinely outproduce a much larger one that has bolted a chat window onto an unchanged process.
The tools were the easy part. The reorganisation was the work.