10A roadmap without dates
AI can speed up the building and record every story as it goes. Deciding what to build, and in what order, stays with the team. I had a roadmap from the start and no forecast, and the reason says something about where the bottleneck moves when code becomes cheap.
Planning happens at several horizons, the product, the release, the sprint and the day, each detailed only as far as knowledge allows, and the first release should be a walking skeleton (Agile Planning). That is how this project was planned. BMAD covers the near horizons well: epics, stories, a status file that says where every story stands, and change proposals when the plan has to move. The product horizon stayed with the team, which here meant me.
The roadmap
The roadmap was a sequence of stages, each ending with a question answered.
- Before the project: what is generative AI worth, what is it worth for building software, and with which method? Three gates, described in the prologue and in chapter 04. A CTO does not take an organisation down a path that turns out to be a dead end, and I did not take myself down one without checking first.
- The walking skeleton. Authentication, configuration, the management and publication of configuration, and the application on a deliberately narrow scope, down to the database, with traceability and auditability built in. Done (chapter 03).
- The first clinical workflows. The core of what clinicians will use.
- A second module, first slice.
- The same module, second slice.
- What comes next, still to be defined.
The near stages are broken down into epics and user stories; the far ones are a line each, and will stay that way until they come closer. My rough estimate is that stages 3 to 5 together are at least five times the size of the walking skeleton, and even that figure may change, because I plan to automate more of the work with tooling that is not yet available.
A roadmap also needs conditions under which it stops, and this one has three.
When the roadmap stops
The roadmap stops if the effort turns out to be more than the project can take on; if Claude and BMAD turn out not to be the right combination after all, which after twelve weeks looks very unlikely; or if there is no business case. Stopping for any of these would be the business case doing its job.
What I did not have
A roadmap says in what order. A forecast says by when, as a range: this slice will most likely be done between these two dates. I did not make one, for a practical reason. The project is self-funded, half-time at most, and my capacity is irregular: some weeks are given to the product, others to everything else. A plan with dates would have been broken every month, and a plan that is always broken stops being read.
The last three weeks matter as much as the others. Only two stories were closed, both built in September and closed once checked live on 16 October, and the project was busy: I was redefining the product's scope for the stages ahead. With AI writing the code, the time spent building shrinks, and the time spent deciding what to build stays what it was.
Where the bottleneck went
Planning a traditional software project means estimating how long developers will take to build things. On this project, building was rarely what limited progress. My time was: deciding what to build, answering the questions the design sessions raised, reading review findings, checking the result live. In twelve weeks, I cannot point to a single week where the code held the project back. Each check is cheap, but together they add up to most of a week. Every review finding to settle and every story to check live takes minutes, and together they take most of my time, which is also why the output can flood the person reading it (chapter 05).
Counting stories still works, then, as long as the count is read as a measure of how many decisions and checks I can get through in a week. And when building is cheap, the uncertainty moves to the parts that are not: understanding a clinical workflow well enough to specify it, and validating that the result is safe. Those will dominate the next stages, more than they dominated the skeleton.
A gate passed
The first stage was meant to answer whether a highly configurable medical device can be configured, approved, published and used from end to end, on foundations built to the standard the device requires. The walking skeleton answered yes (chapter 03). As the articles on this site recommend, the next decision is now on the table, with the evidence needed to make it.
The next stages call for capacity: structured work with hospital practitioners, clinical input and validation that one person cannot provide alone, and more than half of one person's time. Each stage is funded once the previous one has answered its question, and this one has.
That is also why a forecast made today would be of limited use: at half-time, it would mostly measure my availability. The forecast that matters is the one made once the capacity for the next stage is known, and I will keep it out of this article, because when the product could reach hospitals is commercially sensitive.
How the forecast will be made
The data for it exists, because BMAD and version control recorded every story and every week without anyone filling in a timesheet. The method is the one from Agile Planning: size the next slices in stories, divide by the observed range of throughput, and state the result as a range.
The forecast will follow the decision on the next stage, and it will be updated as each slice teaches me how fast it really goes.
For your own project
- Let the AI cover the near horizons, stories, status and change proposals, and keep the product horizon, what to build in what order and why, with the team. Write it down even if no tool asks for it.
- When AI writes the code, measure throughput in decisions and checks, and expect the uncertainty to move from building to understanding and validating.
- Keep the records a forecast needs. With an AI working in a structured method, they are produced as a side effect of the work: every story, its status and its date.