Endgame Viable

Everybody hold your breath, this post is about AI, but it has nothing to do with stealing artwork or data center overcrowding. It’s a practical guide to using agentic AI assistants, something I don’t see written down very often. Usually I only see AI written about as a nebulous malevolent force hiding behind a foggy shroud of distrust.

By way of introduction, and as a way to shoehorn a different draft into this post, I pose the question: Is AI something new, or is AI merely this generation’s version of a profound technological advancement in productivity and convenience:

• The washing machine replacing the need to pound clothes on a rock or whatever.
• Automobiles replacing horses.
• Synthesizers and sequencers replacing musical instruments played by musicians.
• Digital Audio Workstations (computers) replacing analog mixing consoles and tape machines in music production.
• Autotune replacing the need to practice singing.
• Intellisense replacing programming documentation.
• Syntax highlighting replacing the need to format readable source code.
• Electricity replacing gas and oil and steam.
• Plumbing replacing walking down to the river with a bucket.
• Firearms replacing crossbows.
• Spellcheckers replacing the need to memorize vocabulary.
• Calculators replacing the need to learn math.
• Google replacing the need to remember facts.
• Desktop publishing replacing the traditional paste-up process.
• Steam engines replacing manual labor.
• Accounting software replacing paper ledgers.
• Cell phones and tablets replacing desktop computers.
• Touch screens replacing mice.
• Computers and drawing tablets replacing canvases and paint brushes.

My current thinking puts AI in the same category as all of the things on that list. I may be wrong… AI may well be different, but how? It’s not clear to me yet. And so I approach it as something new to learn.

At home, I pretty much live in Google’s Antigravity IDE now, and at work, for software development, Claude Code is interwoven into everything we do. In other words, I use agentic AI assistance for development tasks pretty much every day.

When I say “agentic AI,” by the way, I’m not just talking about asking a chatbot some questions. I’m talking about a chatbot that “thinks.” (Airquotes because it doesn’t actually think.) It’s a chatbot that’s connected to your files and directories. You give it a goal, and it “figures things out” without having to be told what to do at every step. It either writes or runs the tools it needs to get the tasks done, while you sit back and watch.

Below are some prompts that I find myself returning to over and over again, or have used before, to make my life easier.

Code-related prompts

• Review this code to improve code clarity, reduce duplication, reduce complexity, improve ease of maintenance, and update legacy code style. (Often good to do a pass or two after AI itself writes some code.)
• Review this code for leaks of configuration secrets (before committing to git).
• Document step-by-step instructions for how this code is deployed to a production environment.
• Describe and diagram the architecture and design choices of this code (useful when approaching a new code base, something I have to do at work all the time).
• Write unit tests for the new code I just added (I’m still developing strategies for prompting AI to write good unit tests, it usually brute-forces them; some folks at work are experimenting with test-driven AI development, where they get the AI to write the unit tests first before writing the code, which is a good idea I want to try someday).
• Review this code for tightly coupled objects and suggest a plan to decouple them.
• Build a localstack test environment for this aws code.
• Look at these log errors and find the root cause of the error and suggest a solution.
• Fetch the log lines related to request ID x and trace the end to end operation of the request.

Not-code related prompts

• Search these hundreds of photo files to find instances where there’s a break in the sequential numbering and rename any incorrect files, and update all references to those renamed files.
• Create a static web site generator to publish these hundreds of photos of an old World War 1 diary and the transcripts of each page, converting each RAW photo to a small web-friendly webp format.
• Import the blog posts saved in this WordPress backup XML file as Markdown files for Hugo.
• Search these hundreds of Markdown blog post files for broken image links and correct them, using this directory of images downloaded from WordPress to match the missing images. And move those WordPress images to my S3 bucket directory, resize, and convert them to webp.
• Connect to aws and download the raw access logs for my blog, then generate an HTML summary report of statistics. (Real-life example here.)
• Create a new Hugo blog theme with a more modern web style that supports dark and light mode.
• Write a tool to use an api to transcribe the audio commentary track of these hundreds of videos as SRT files, and create a summary description of each video from each transcript. (I actually wrote most of that functionality before Antigravity, but this is how I’d do it now.)
• Give me the steps to migrate my gotosocial instance and postgresql database from an unraid server on my local network to a VPS in the cloud.

If you don’t know, what happens with Claude or Antigravity (and I assume all others) is that after entering a prompt like the ones above, the AI will generate a “plan,” which is a long document that describes what it’s going to do in step-by-step detail. It may ask you to clarify things. You look over the plan, find things you want to change or improve (there’s always something), tell the AI what to do different, review the new plan, repeat, and eventually approve the plan and watch it go.

In my mind, I assume everyone who works in technology has already figured these kinds of prompts out, and probably better ones, and doesn’t need to be told any of this. But on the off chance that somebody out there still doesn’t know, or maybe hasn’t had a chance to approach agentic AI yet, this may spark some ideas of what is possible. You always have to start somewhere with new technology.

Here is an example session with Antigravity about migrating gotosocial that I copied into a page, if you want to see the conversational process in action.

With those prompts, the agentic AI is usually being directed to work on files (e.g. source code, but it could just as easily be Markdown blog files or picture files or video files) in the current directory or git repository.

I’ve done all of the things on those lists, so they are reflective of the kinds of things I’m using AI for.

Those are just samples to give you ideas. You will undoubtedly need to converse with the AI to get all the details exactly right for your situation. (See the gotosocial example above.)

Honestly the list of uses just goes on and on and on. I use it all the time. I can barely remember what it was like to have to remember or lookup complex command or programming language syntax to do all those things by hand.

Just recently I used it to decode a MIDI timing sync audio signal from a recording from 1995, and generate timing information for a modern REAPER project.

I don’t know what the future of AI holds, legally or ethically. All I know is, like many technical advancements of the past we now take for granted, like the IDE, or syntax highlighting, or Intellisense, it sure would suck to have to go back to doing general development tasks the hard and slow way.