My evolving working practice, with Claude, Guru and other assistants.
Claude was my first AI collaborator and did a great deal of the work with me. That deserves a direct acknowledgment. The software, writing and creative projects here did not come from me typing every line alone. They came from a process of asking, inspecting, rejecting, correcting and trying again.
As the number of projects grew, the recurring problem became continuity. An assistant could sound certain about what was finished while a check was still missing. A correction could disappear into a conversation. Work could be technically complete and still be difficult to use.
Guru, Homebase, dev-teamplus, Command Center and the quality-checking tools address different parts of that problem. I am interested in how the work is assigned, how it is recovered after an interruption, and what evidence is required before I call it finished. The number of agents involved is not a measure of success.
This website supplied a small, public example. I rejected two visual directions before asking for sections that actually represented their subjects. After launch, Claude’s review caught weak mobile navigation, an awkward CNC image and stories that were too vague. Working links and successful publication had not established that the presentation was good enough.
The correction was a better question: can a visitor understand the work and see a reason to trust my judgment? That led to another design pass and more concrete accounts of failures and decisions. It also required challenging parts of the review rather than accepting every confident diagnosis.
I keep the reusable machinery private, but the lesson is public. AI makes it possible to produce more work quickly. It does not relieve me of deciding what should be built, checking what actually happened, or putting my name behind the result. That is the part I bring to a consulting engagement.
Working with ChatGPT
ChatGPT, through Codex, has become another collaborator in that process. On this site it helped review the project inventory, develop the stories, create the illustrated presentation, build the editing interface and publish the result on my domain. It also made mistakes: the first designs missed what I wanted, and the early writing protected the implementation so carefully that too much substance disappeared.
Claude’s review gave us another perspective. I asked ChatGPT to examine it, challenge unsupported conclusions and make the useful changes. That exchange is how I want to work: credit the contribution, keep the evidence visible and revise the result when it falls short. Neither assistant gets a pass because its answer sounds confident. My decisions and the published outcome remain my responsibility.