The Projects AI Makes Possible

A CD-ROM, coding laptop, sketchbook and tablet almanac on a home workbench.

My family was playing an old Carmen Sandiego game, and I wondered whether we could build our own tablet-friendly almanac to help crack the clues. Somewhere on the original CD-ROM were the facts, pictures, and maps. Could an AI agent help me get them out?

Previously, I probably would have filed that idea under “someday.” This time, I could describe what I wanted and let an agent start exploring, without committing my evenings to an idea that might go nowhere. We followed false leads, spotted pictures emerging from scrambled pixels, and left experiments running overnight. It was fun to see how far we could take it.

These eight essays follow the project from that first question to an almanac we could actually play with, including the mistakes, surprises, and moments when I had to rethink what I was asking the agent to do.

Read from Part 1 for the whole story, or choose an essay below. Each can be read on its own.

  1. Part 1: The Project I Would Never Have Started Alone

    AI Tooling lowers the cost of experimentation, encouraging big swings on speculative projects. With agents handling the busywork, it's easy to try something that may or may not work, and cost-effective to attempt potential dead ends that would otherwise make some projects hard to justify.

  2. Part 2: Beginning To Build With AI – And Pivoting To Prioritizing Review Tooling

    While decoding the archive, I quickly realized we needed some kind of tooling to review that was more ergonomic than simply chat messages. It was exciting when the agent served up formatted, organized web pages to let me quickly see country by country what had failed, what had succeeded, and how other experiments were going.

  3. Part 3: The Agent Explored Different Paths. I Chose Which To Follow.

    A screenshot gave the agent a direction, but human judgment spotted the faint resemblance that helped us move from failed renders to a working decoder.

  4. Part 4: Learning to Trust an AI Agent With the Goal, Not Dictate Every Step

    The recovered maps were too small to reuse. An agent rebuilt thirteen modern equivalents overnight, preserving the old game's regional logic without preserving its pixels.

  5. Part 5: The Agent Built What I Asked, But Not What I Meant

    Unclear requirements led the agent to produce a polished AI rewrite that sounded complete while dropping the details that mattered. Once again, building a system to validate the output was as important as the system and process generating the output.

  6. Part 6: AI Solved the Problem I Pointed It At. I Had Aimed Too Narrowly.

    One failed search during gameplay exposed a swath of data we had initially missed. This led us to greatly expand our source material and audit 5,441 records, ultimately resolving the missing entries and improving search.

  7. Part 7: Strategic Pruning: Deciding What Belongs in a Minimum Viable Product

    The low cost of exploration and addition led to a quickly expanding project. The final step was going in the opposite direction - reducing what made it into the final product, allowing the almanac to fade into the background and be a small piece of the overall detective experience.

  8. Part 8: Reflections on Collaborating With an AI Agent

    Building a private almanac with an AI agent gave me a chance to work differently: let it explore, inspect what came back, and help it find the next direction without needing to work through every implementation detail myself.