For the past two days, I worked on pivoting Papersy into a human-agent workspace using SDD, or Spec-driven development. This is for an exercise to get hired at some place, though I highly doubt they will answer. And I must say, I have mixed feelings about this process.
For starters, the spec -> plan way is good. I used Opus High to do SDD so it had to be good. The way the coding agent planned everything, with me going back and forth with it (using grill-me skill. Thanks, Matt Pocock) to adjust the documents. I read both SPEC.md and PLAN.md, because they were easier to read and didn't go fully into the nitty-gritty details like TASKS.md, which was why I skimmed it.
After that came the implementation phase. Claude Code used Sonnet Medium for that phase. After a few hours (I'm not exaggerating, that was just for frontend, backend, data model, OpenRouter, RAG), it finished coding, with unit tests and e2e tests (because I told it to do so). It also ran a code-review (for some reason couldn't be activated by the agent, had to activate manually) using Opus High, along with that High effort model fixing all errors. Later, Sonnet Medium was used to run all tests. Then, Haiku ran through the code and documented the pivot. All that cost $100 (a 5-hour session + extra $40 usage credits). During the procedure, Claude Code closely followed the code and tests against the accepted criteria, which is good.
However, I tested it the next morning. And for the most part, it worked fine. But I still had to straighten out some styling issues. And one particular error relating to PDF upload from markitdown-ts, which I discovered through AI debugging, was built from pdf-parse. Turns out after 5 hours of wrestling with Claude about the way to fix it, a trip down Google Search turned up that troubleshooting doc that told me exactly how to fix it. Imagine having up to 10 commits of logging commands to Vercel, just to catch this one particular bug. I guess with a more sophisticated AI self-healing software, that could have been avoided. But until that day comes, I still have a job to do.
That's what I have been thinking after using this SDD and created this AI workspace pivot. Now, a coding agent can spawn an almost perfect app for a price lower than my salary. But without expertise, the app would never work correctly. You can have all the knowledge at your fingertips, and you still won't know anything due to lack of expertise. I can't claim to be a lawyer even with an AI agent to help me find out about laws. Or a doctor by watching House, M.D. I can be more informed, but the thinking and decisions have to belong to the expert.
The world is changing, no doubt about it. I just don't know to what. Let's hope I don't visit a doctor who "vibes" his way through medical school, or a lawyer who "vibes" his case (that has happened more times than you think). I have enough problems with everyone "vibing" and selling their own solutions on LinkedIn, Facebook, GitHub, whether they are software solutions or a bunch of skills packaged as a Claude plugin. Nowadays, I tend to only use and learn a new skill, library, technique when it serves an actual purpose.