How I Use AI for Software Development

TLDR; I’ve been using AI. I’ve changed my approach a lot as I’ve gained experience. AI has allowed me to experiment with more tools and process, and build more, than I could possibly have done without it. I’m experimenting to find new ways to use it better.

My use of AI evolves. Every day.

I use AI. I learn what worked and what didn’t. I change the way I work. How I use AI is not fixed.

This post is a snapshot summary of what I’m currently doing and I’ll provide links to experiments and results of those experiments later. I’ll also point to a prompt I’ve been using to interact with an application and find defects. And I’ve listed some of the AI tools I’m gaining most value from.

I hope to create more posts that describe the path I used to get the the current way I use AI, lessons learned, and how I’m using AI in more detail in future posts.

This post was written by a human. Based on the human’s experiences of using AI to get things done.

I’m The Human. Me. I wrote this.

How I used AI this Morning

I don’t think I have time to list all the activities, and products, and ways, that I’ve used AI over the last 6 - 8 months so it is probably easier to start with a snapshot of how I used it this morning 26th June 2026.

This morning I used AI for:

• work on Project A
  • automatically reviewing a PR
  • respond to and fix PR request comments
  • interactively use Playwright to explore changes related to a PR and create some defects
  • write code to fix the defects raised by the earlier AI process
  • auto create videos when I asked AI to check if a defect was replicable
• work on Project B
  • fix some feature nuances and make the app work the way I want to
  • investigate why downloads for some content were not working as well as they should
• personal development
  • summarize 20 new podcast episodes so I can review them and see if anything is useful
  • search the web and summarize news stories on a variety of topics

My role on Project A this morning - Product Owner & Manager

My role on Project A? On Project A today I was the Product Owner and Engineering Manager.

• I triaged the PR comments and closed any I thought were not relevant prior to the AI review
• I triaged the defects and closed any that I thought were not relevant and added ‘fix it this way’ comments for those that could be approached ambiguously
• created new stories to expand functionality and move the project forwards
• adjusted the way the AI interactive review process works and I’ll see if it is improved later as a result
• this afternoon I hope to test and review the changes (But it’s so hot! And I’m writing this post.)

Project A - this morning was primarily interactive review coverage for work that yesterday I was the Staff Engineer. Yesterday, I reviewed the code and architecture decisions, and specified the work since it was a technical re-architecture story.

Product A is AnyWayData.com and I’m approaching this as a professional development project and trying to learn how AI can work for a Software Development Team.

My Role on Project B this morning - User

My role on Project B? On Project B - I was the user. The product is fairly mature now and is in daily production usage, by me. I use the product, and when I see something that can be improved, I tap the AI on the shoulder and say - can you improve this… then it does as I continue working… then I test the improvement and keep going.

Product B is an internal tool, but it is the basis for the podcast summaries on:

• TesterHQ.com
• BoomStash.com
• pdf summary briefings on Patreon

This is ’less professional team’ and more ‘solo dev’ but has automated coverage, code reviews, CI, database, CDN, etc. etc. I interact with AI differently when building these products.

Multiple Roles

How I use AI varies from day to day and I’m still learning what it is capable of.

I have to adopt all the different Software Development roles when I develop with AI.

I do not believe we will one day all just be Project Owners. It seems more likely that to get the most out of AI for Software Development we will need to learn how to do all of the processes in Software Development to some extent.

Teams that adopt AI can have multiple people and each person will adopt multiple ‘senior’ roles, while the AI takes on the junior role.

The smaller you make the teams. The more roles each team member has to learn how to do.

The faster you move, the more you have to trust the AI processes you put in place. So only move as fast as you trust the process.

As the Product Owner for the AI:

• I come up with the ’next thing to do’
• I write the acceptance criteria
• I plan the scope of the product

I have to act as Senior Engineer:

• I’m responsible for the architecture and coding principles that are used.
• I’m pushing for more tooling to help improve quality and speed of delivery
• I’m pushing for a better build process and better environment release process
• I review the code because the AI cheats and takes short cuts when it writes code and if I don’t spot those in a review it impacts long term development

I have to act as Tester because:

• the AI is getting better at finding defects by performing focused exploration on the changes made but still not good at managing coverage
• the AI still operates at a fairly junior level and has biases in its approach
• the AI often makes mistakes when communicating a defect and I have to identify the root cause
• the AI can simulate some of the activities that a Tester performs, but it’s an AI, not a Tester.

I have to act as Senior Automator because:

• AI writes horrible automated execution code without guidance and review
• I have to own the standards that we automate to

I have to act as Manager to prioritise the build work because:

• The AI doesn’t know how to prioritise
• The AI doesn’t know how to interpret nuance

When I say “the AI doesn’t know how to…” I’m really saying “I haven’t found a way to delegate this to the AI such that I would trust the output without getting involved.”

I’m still learning how to work with AI and I think I’m learning how to use it better every day.

Can AI be a team sport?

When I first started working with AI, and when other teams around me were working with AI, it didn’t feel like it would fit in well to a team process.

• People using AI had a hard time getting AI generated PRs through a Human Review.
• It was hard to reign in AI and keep it focussed on a single section of code so PRs would expand and merge conflicts were plentiful.
• AI seemed to rely on specs and docs in the code and didn’t integrate with any Human Tooling.

Now, having used AI for some time. I’m trying to add more ’team’ structure to the process.

This means I’m forcing AI to adapt to a human communication process more than me adapt to what it needs.

My current approach to AI on AnyWayData is one that I think can be used in teams.

• Create Github Issues for the work to do with acceptance criteria
• ‘plan’ with the AI around that issue - AI tooling can read the Github issue directly so I don’t need to cut and paste into the system
• AI can write the code
• AI can write supporting automated coverage
• AI comments on Github with Summaries and artifacts from the work done
• Create PRs for changes
• AI Tooling automatically reviews the PRs and leaves comments
• AI and Humans can review the comments and triage them
• AI can read the PR and resolve comments
• Frequent commits mean we can have multiple restore points or commits to cherry pick from
• AI can be pointed at an issue and a PR to provide a narrower context to interact with the application and identify defects in the running system, not just code review

As a human, I’m involved in all those steps. And I’m moving some of my review work into Github, just like I would do on a human team. I’m trying to learn how to fit the AI into the process so that Human work remains central for evaluation and decision making.

Can Humans keep up?

I see a lot of tool vendor statements that AI is now producing code so fast that humans can’t keep up.

And therefore we need to have a new breed of tools that do other things fast.

Really?

If I can’t keep up with the code creation, how will I find time to also keep up with the AI generated output from the other tools?

And if the other tools can do things fast, won’t that allow the code creation to move even faster? A continually accelerating cycle that will eventually move so quickly the output will evolve faster than anyone can use. A process limited only by the heat sinks on our machines and the speed of our deployment process.

Humans can keep up.

If the humans make the AI work at the speed and level of quality that they can handle.

We can choose to create with AI in ways that we can keep up.

There has to be a better marketing tag line.

I am available for marketing consultancy, contact me here.

If Humans can keep up then we aren’t maxing out our use of tokens and are wasting money!

We might then argue that if puny humans can keep up, then we aren’t making the most use of AI and are letting tokens go to waste.

The response is not - then everything must go fast - it could be - “you’re paying $200 a month, that’s less than $10 a working day. and if you are getting more than $10 a day in value from the AI tooling, then you’re doing fine”.

Decide on the monetary return in value before deciding you are wasting money. Experiment to increase the return in value and improve gradually.

Perhaps use those extra tokens that you are ‘wasting’ for side investigations that you don’t have to keep up with?

Here are some side-investigations I performed this week. Because I wanted to and not just to ‘prevent tokens going to waste’.

• AI evaluated an ‘open source’ test tool that I’d seen mentioned while researching things I heard on a podcast. Tried to get it working. Amended code to bypass some hard environment dependencies that limited its use as open source. Provided in depth analysis of its features and how it worked.
• AI evaluated 5 or 6 static analysis tools that I don’t have on the project. Downloaded, installed, used, timed the performance, compared the features, described the impact of adopting them, ruled out those that wouldn’t work.
• AI investigated and implemented CDN caching for one of my apps to reduce hosting cost.

I now have decided not to spend time on the ‘open source’ test tool. And have made some decisions about tooling migrations that I will prioritise next week. And I think I’ve saved a whole $4.93 in hosting costs per month.

There are ‘other things’ the AI can do, in parallel, that don’t force the humans to go faster.

Focus and Experiment at the same time

Our focus on projects has often been - prioritise, and focus on 1 or 2 achievable clear goals, because we can’t do everything.

Now we can do more.

Experiment.

Have the AI tooling do things in parallel that we would never get the chance to do. While we maintain a focus on 1 or 2 achievable clear goals that we can pursue at a sustainable pace.

Use the extra tokens to experiment:

• Some of those experiments will probably end up being throwaway.
• Some you will forget to look at - I can see in my Codex threads that I have an experiment from a month ago that I have yet to try (no, I did not just inadvertently provide evidence that humans can’t keep up).
• Some you will prioritise to take forward at a sustainable pace.

Some of those experiments will also be on how far we can push AI and our process.

My use of AI is evolving. I might even learn to trust the more Agentic styles of development. I currently don’t. But I also haven’t had time to experiment with them because I’m getting things done with my current evolving approach.

Where can I see some evidence and output?

I’ve been using AI to expand AnyWayData.com and I’m pushing AI to see how much I can use AI, how effective AI is for development and experimentation, and not just - how fast can I expand the functionality.

Prior to May 2026 AnyWayData was a legacy human written project with no real automated execution coverage.

I initially used AI to finish migrating the tool’s data grid from AG Grid to Tabulator. I can see from commit history that I had started work on that in June 2025 and intermittently worked on it until December 2025. Then stopped. I remember it only took a weekend to finish the migration when I started back working on it with AI - of course, I, The ‘Puny’ Human, had done the hard work to architect the abstraction layer required to make the AI step easy!

And then I started expanding AnyWayData to experiment with the process and technology of AI and learn how to build more professionally with AI.

Using AI I’ve added:

• Storybook - to help refactor to UI components and introduce a lower level of interaction coverage.
• Playwright for automated API and deployed UI coverage
• Dockerisation of the MCP, and API
• An auto-deployed test environment
  • https://eviltester.github.io/grid-table-editor/
• Improved Github pipelines
• Features that I had been experimenting with but never finished e.g n-wise data generation
• Experiments with WebMCP and in browser AI APIs
• An experimental AI interactive review process which reports defects
  • https://github.com/eviltester/anywaydata-ai-testing
• and more…

Looking through Github commits I’m surprised that I only started using AI on the project extensively at the start of May (end of April) 2026. Because I’ve done more in that time, than I could possibly have done on my own without the support of AI. I didn’t realise I had expanded the functionality quite so much and quite so quickly - that’s a pretty good secondary gain.

Over that time I’ve used:

• all the free AI chat systems
• OpenCode
• Continue.dev
• Co-Pilot
• Claude Code
• Codex
• OpenRouter

I currently use OpenCode (for manually triggered reviews) and Codex (for everything else). And for the first half of the month I use Co-Pilot credits to review PRs until they run out (I suspect co-pilot might not be in my tool list much longer).

I use OpenRouter as my AI API marketplace, rather than run all models locally.

I also use the open source plans of the following tools for AI Automated PR code reviews:

• CodeRabbit
• Greptile

You can see review comments from both of the above on my recent PRs.

I’m also using the project to try and experiment with other tooling.

I’ve even tried a few commercial AI test execution tools on the project but the trials are often too limited or I just can’t pull value from them. But since the tools evolve quickly I look forward to trying more commercial tools in the future and revisiting those that I have tried earlier.

I’m trying to make as much of this work as public as I can, and as visible as I can, to make it easier to pull lessons from.

• https://github.com/eviltester/grid-table-editor

If you look at the PRs in the project then you’ll see the comments generated by the AI review tools that I’ve been using.

And I think, if you looked through the most recent closed issues and PRs you can get a feel for how I’m trying to use AI in a process that could work in a team environment.

I haven’t really been blogging about this.

I’ve been describing progress in Patreon posts because I’ve been experimenting as much as I can, rather than reporting on work in progress that was too early to comment on properly.

Can AI Test Stuff?

I’ve been evolving a prompt that I have found very useful as part of my testing process over at:

• https://github.com/eviltester/anywaydata-ai-testing

It’s just a prompt. It is not a new revolutionary AI Testing framework.

It was easy to install:

• configure Playwright MCP
• or Chrome Dev Tools MCP
• paste in prompt to Codex

I wouldn’t call it testing. My mind frames it as an Automated Interactive AI Review.

I’m adding the output from the process in the repo to make it easy to see the output and you can review to see if it might add value. Hint: it’s in the main readme.md

In the prompt I use the language associated with testing because I’m trying to set the context for the AI such that it communicates and operates as a simulation of a Tester and I’ll see how much value I can extract from it.

I will say that in the last few days, this has generated more valuable defects for me than any of the AI Test Tools I’ve used.

Have a look through the repo, experiment with the approach and see if it adds value.

Can AI Test Stuff? On its own? No.

Can we use AI as part of our test process? Yes. Hopefully the repo above provides a simple example of an experiment to try as part of your test process and you can use it in any tool.

Have a look through the AnyWayData code:

• https://github.com/eviltester/grid-table-editor

And you’ll see I’m using AI to generate a lot of execution coverage, using multiple tools.

I use Page Objects to model the UI. I find that works better when the AI tooling expands coverage.

I did try and use the Playwright MCP to generate execution code but it was… not acceptable to me.

Hopefully there are some jumping off points here to see some examples to learn from and some of the experiments I’ve been running.

How to Learn AI?

The current approach that I’m using now feels very different to the experiments I was running when I started to learn AI.

I documented a few of those experiments here:

• https://github.com/eviltester/ai-supported-testing-experiments

So the way to learn AI is:

• pick an AI tool
• use it to try and do something
• look at what didn’t work
• look at what did work
• try to do the thing again, but better

Experiment and iterate.

And… the more you can share what you are doing, the more everyone else can learn as well.

Don’t be afraid to pursue a different direction to everyone else.

Your natural approach will probably be completely different to someone else’s. You’ll do experiments that other people won’t.

I’m learning from the excellent work that Dragan Spiridonov is doing, and the comprehensive information he is sharing.

But… I’m experimenting with approaches that fit my natural approach, and seeing which add value to my process.

Everyone working with AI at the moment is learning.

Start with the first experiment that makes sense to you and keep going.