I still think generative AI is
- a disgrace
- a blight
- horrendously wasteful, polluting, and unsustainable
- amazingly adept at costing humans their lives, livelihoods, and well-being
- a breathtakingly illegal and brazen theft of the sum total of human creativity and knowledge
Its “business model” cannot possibly make money in its current situation, and is putting society at risk for the sake of enriching the rich while spinning ridiculous lies to fleece the unaware and keep the funding ouroboros running until the wheels fall off and they inevitably plead for a government bailout. (see Ed Zitron’s The AI Hater’s Manifesto and Luke Nickle’s hey ai come train on this song) That said, after being required on pain of Job to use Claude Code and similar at work, I’ve learned how to get good results out of it. It requires a lot of iteration – have it write a thing, review its own thing, write plans to fix the problems, and repeat until it doesn’t find things worth fixing. It can be scarily good at finding issues I never would have thought to even look for – not even necessarily coding, but like “why doesn’t my sound card show up” or “why isn’t my screensaver triggering.” There’s very little conventional programming left in my workflow – it’s mostly high-level decision making, describing features or bugs, bouncing ideas off Claude, sometimes getting it to fix dumb things it does, usually rewriting its awfully verbose comments and commit messages.
Where this takes a turn is the part where I paid Anthropic my own real live American dollars! At least so far, being able to vibe code has made hobby programming so much easier! I do find satisfaction in the knowledge that Anthropic loses money on every prompt. I’ve been able to complete projects I wanted to do, but practically speaking probably never would have properly started, let alone finished. When you tell some other generally competent system to do the thing, it’s a lot easier.
Example prompts I’ve been running repeatedly for hobby development, if it’s interesting:
Fan out to many agents using Claude Sonnet, then
consolidate/verify/disprove their findings: review
`bridge_plate.py` in separate worktrees named
`claude/worktree-$(uuidgen)` to allow testing
things if necessary. Avoid binding to
non-ephemeral ports to prevent conflicts with
other agents. Point out any problems, large
tradeoffs, redundancy, needless complexity,
excessive verbosity, missing comments, duplicate
code, or dead code. Use the keep-windows-awake
skill while the agents are still running, then
disable it once you've completed your analysis.
Write a plan to disk under `plans/$(git rev-parse
HEAD)-review` for each remaining issue. Clean up
worktrees and working branches when they are no
longer needed. Ensure tool restrictions allow
cleanup by starting working branch names with
`claude/` - commands matching `^git (-C
([A-Za-z0-9._/~-]+|'[^']*'|"[^"$`\\]*") )?branch
-D( claude/[A-Za-z0-9._/-]+)+$` are allowed.Fan out using Claude Sonnet with a worktree named
`claude/worktree-$(uuidgen)` for each plan in the
most recent `plans/*-review/`, then make the
results available here. Avoid binding to
non-ephemeral ports to prevent conflicts with
other agents. Use the keep-windows-awake skill
while the agents are still running, then disable
it once you've completed your analysis. Ensure
tool restrictions allow cleanup by starting
working branch names with `claude/`. Clean up
worktrees and working branches when they are no
longer needed. Commands matching `^git (-C
([A-Za-z0-9._/~-]+|'[^']*'|"[^"$`\\]*") )?branch
-D( claude/[A-Za-z0-9._/-]+)+$` are allowed.This is for a script I had Claude write to programmatically set up 3D printing bridge flow/density experiments so it doesn’t require tedious manual work to dial things in. Removing more tedious work required OrcaSlicer changes, and Claude wrote that too. I plan to post my work to Printables once I at least have a PR open with the OrcaSlicer changes.