Claude and agentic building
When the model can read your tree, run commands, and hand you diffs, the shape of work changes. Not magic. Closer to a very fast teammate who still needs you to steer.
TL;DR
- For me “agentic” means a tight loop: plan, touch files or run commands, look at output, fix.
- Claude is strongest when it sees your real errors and the task has a clear finish line.
- Boundaries, regressions, secrets, and “should we even build this” stay on you.
What I mean when I say agentic
Not a manifesto. In projects where I leaned on Claude and similar setups, it meant the model was not trapped in a chat bubble. It could read files, propose edits, run checks, fail, try again. The loop matters. Short cycles where the machine does tedious work and I stay on steering, constraints, and what counts as done.
Different from “paste this function.” More like pairing with someone who does not get tired of boilerplate but still needs you to say when the design is off.
Where it actually helped
Scaffolding and migrations. New folders, wiring config, repetitive refactors. Work that is easy to mess up because it is boring. A flow that can touch many files and show a diff is useful if you review like you would a junior’s PR.
Closing the loop on errors. Stack traces and red tests are structured. When the model sees the same terminal output you see, you stop retyping logs into another window. Shared state speeds things up.
Getting oriented. I still treat explanations as guesses across versions, but a walkthrough of “what this module thinks it is doing” can shrink ramp time before I lock an architecture bet.
Where it sounds sure and is not
Confidence. The voice is steady. The world is not. Sometimes the bug is not syntax. It is a pretty abstraction that couples things you meant to keep apart, or hand waves past edge cases. That is why I care more about tests, types, and thin module edges when I am moving fast, not less.
Drift across sessions. Without habits you stack “just one more fix” layers. I like checkpoints: commits that mean something, flags, simple rules like “this service does not call that one directly.” Otherwise you get speed today and archaeology next month.
Security. Anything that can run commands gets the same paranoia I would use for a new hire with shell access. Tight permissions, review before destructive stuff, no fantasy about full autonomy.
Keeping agents from owning the shape
I want future me, on a bad Tuesday, to still read the repo. That usually means small modules with obvious contracts, at least one test that would have caught the clever merge, and a short note when a decision is not obvious. README paragraph counts. So does a design note in the repo. Chat logs do not count as documentation.
The model can draft those. Requiring them is still my call. That instinct is not something I plan to outsource.
Claude does not replace product intent
Ties to the other piece I wrote about what to build. Claude can help you run hard in a direction. It will not, by itself, tell you if that direction was worth support load, maintenance, or the opportunity cost of everything you skipped.
Split I am ok with: tools for execution and exploration, me for problem choice, tradeoffs, and the overall shape. The sessions I liked were not the loudest demos. They were the ones with tight loops, reviewable diffs, and a goal written down before anything started cheering.