Just before I left my last role we were debating the format for stories and requirements and anything else for the engineers to use as input into the AI “factory” that would build something from the input.
When I first started using AI as a one off task machine, I would spend quite a while creating the right prompt, but that goes to another level when you are building a team of agents to build a complex system. Doing so is a complex task and very hard so I will completely ignore it.
The debate we were having started as a conversation about “Will AI mean we should get everything right up front and give it to the coding agents, like waterfall?”. I argued for lots of iterations, my friend agreed but felt that being right at the start was better than not giving the agent everything. So some planning and some context building could turn into some real work.
The conversation moved to stage gates and we both agreed that we should put all stage gates up for review – culling them wantonly or reengineering them massively. So I argued for no more “discovery complete” or “design complete” and he argued for “Build good standards and context” and “Use AI at every stage so we can still do a design we are happy with”.
Much excitement and experiment ensued. I imagine the experiments continue and that the “AI Factory” of agents has evolved even in the short time since I was there.
But one unresolved point was building documents for humans during the progression of the idea from insights, to commitments to breaking the work down. It is fine to say “AI can do lots of stuff and we need a human in the loop”.
But if humans are in the loop and we are iterating, then the original idea is evolving as we test some prototype, or even demo a real thing.
When I learned design thinking one thing that jumped out was that it was not just “thinking about design” it was “using the design process to design our thinking”.
So we want people to learn during the process and we need people to align on what we are doing. Gemini can make great presentations and Claude can explain what we are doing in a lot of (potentially correct) detail.
But back when I was a waterfall project manager, people used to assume they knew what was going on because they were at the start of the project and “knew” the original intent and assumptions. Or at least they knew the hypothesis and the known assumptions.
But a big thing was traceability – if we changed the implementation then we should communicate if that changed the design people had seen. If we changed the design then we should communicate that we are changing requirements or even the original problem definition of our whole mission.
Debates around “scope creep” or “MVP for evil” were often about mismatched assumptions about what we thought everyone already knew.
If AI can build and test a complex product in 3 days, then we could “fail in 3 days” and course correct. The cost of misaligned assumptions is pretty small. But that is IF we could build something production ready and maintainable in 3 days and IF stakeholders came to see and test the outcome throughilly.
Anyway – the small topic we discussed in the bigger picture was how AI could handle changing assumptions and expectations.
Certainly it can do a great job of consolidating what is happening and comparing the old and the new. But at some point humans still need to read/interact at a level where they understand things and not at a level where they are aware of the information they could read again.
So – my view was that we should not create a series of documents for the human. We should create a source of truth for the initiative/product/improvement. The source of truth could be a markdown file or a human readable library. Then we keep the source of truth up to date and rely on a skill to reproduce all the documents for humans, together with humans communicating and demonstrating what was really happening.
At the time I left though, every other human in the office wanted to have content consumed by humans rather than an AI to create content on demand. Of course AI would have told us we were all super smart and correct :) .
But looking for a job brought me back to a part of this conversation. AI is now a reality in the job market, both for the candidate and the one hiring someone for a role.
I use AI to review and rate roles before applying. It has really helped me avoid wasting time on things I roles I would not be a good fit for and also gaps in my profile that I can actually fill with real examples.
I also use Co-work now to product my first draft of a cover letter and resume.
I have heard that Anthropic will now watermark my documents and so the recruiter will know Claude was complicit in what I am saying. But that does not worry me because in fact I edit the documents AND they are based on my original specifications. More importantly I am confident that if I get asked about anything in them I can remember what it said and back up what is in there.
Claude and Gemini have both been “creative” in explaining my experience, but the resume that goes out will align with what is in LinkedIn and what someone I worked with will say and what I have really done.
This is partly because I am diligent in wording and editing my professional output. While I am a relaxed person, those who know me also know that I take professional writing seriously whether hand drawn scribbles on a whiteboard or documents I want people to ready thoroughly. Its also because I am a subject matter expert on my history and background.
But I am not a subject matter expert in what the screening AI will do with what I provide. I might say I am good at manageing stakeholders and for all I know the AI translates that as good at aging stakeholders or aging wine.
Presumably in the roles I am likely to apply for there is little need for wine making and less for aging people through harassing them or using some really unhelpful aging tech. Which of course means that a good screeing machine would ditech my resume while a human would human would make sense of it.
I hear lots of “horror” stories of algorithms doing strange things like that. Of course they are likely exaggerated and generally urban myth (internet myth?).
But still – I write a cover letter for a human to read and I rewrite or delete a lot of what uncle claude puts in it. This is both because it is not in my voice and because it is stupid wording that I would not want to read.
But I have a suspicion that the initial audience for my letter is a robot and that maybe Claude understands their “persona” better than I do. Which means what looks stupid to me (the author/editor) is actually written really well for the reader (Hire/delete bot17).
Since the essence of writing is to communicate to the reader, does that make me a bad writer after all? Does it mean I should get Claude to send a letter to HireBot and then if they agree then connect me with a human and let me write an introduction and resume for the human?
Worse – If I craft something with the support of an algorithm and then send it toward a human and it is first caught and summarised by an algorithm, is it really efficient.
I think whether we are writing things for each other or providing context and specs to a factory of engineering agents, the same question is going to come up.
I know when I am writing for Claude and when I am writing a blog to be read by humans. But do I need to relearn the art of writing and editing content when the audience is bot enabled humans, or humans using AI to understand what I am using AI to say?
For now I will direct Claude and write for humans. But am I missing something important here?
James King 