(2025-10-09) Vincent Superpowers How Im Using Coding Agents In October2025

Jesse Vincent: Superpowers: How I'm using coding agents in October 2025. It feels like it was just a couple days ago that I wrote up "How I'm using coding agents in September, 2025".

I've spent the past couple of weeks working on a set of tools to better extract and systematize my processes and to help better steer my agentic buddy

If you want to stop reading and play with my new toys, they're self-driving enough that you can. You'll need Claude Code 2.0.13 or so. Fire it up and then run:
/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplace

After you quit and restart claude, you'll see a new injected prompt:

You have Superpowers.

The coding workflow

It also bakes in the brainstorm -> plan -> implement workflow I've already written about. The biggest change is that you no longer need to run a command or paste in a prompt. If Claude thinks you're trying to start a project or task, it should default into talking through a plan with you before it starts down the path of implementation.

After you're done brainstorming, if you're in a git repo it automatically creates a worktree for the project and changes to that directory.

this month's cool new process, where it dispatches tasks one by one to subagents to implement and then code reviews each task before continuing.

*Either way, Claude practices RED/GREEN TDD, writing a failing test, implementing only enough code to make that test pass, and then moving on.

At the end of the implementation process, Claude will now offer to make a GitHub pull request, merge the worktree back to the source branch locally, or just stop.*

But none of that is the interesting part.

Claude Skills are the interesting part.

Skills are what give your agents Superpowers.

I started to see things that looked a lot like skills everywhere.

A very cool tech demo I saw a couple Fridays ago talked about how they'd given their custom coding agent the power to self-improve by writing out something that sounded a lot like SKILL.md files. Sam Schillace wrote about that demo here.

Sam and Brian Krabach are a couple of the folks behind Microsoft Amplifier, an amazing integrated development framework that uses this same pattern of a coding agent that improves itself by writing out markdown docs and writing tools for itself

One of the first skills I taught Superpowers was How to create skills. That has meant that when I wanted to do something like add git worktree workflows to Superpowers, it was a matter of describing how I wanted the workflows to go...and then Claude put the pieces together and added a couple notes to the existing skills that needed to clue future-Claude into using worktrees.

I haven't published all the skills Claude and I have built

telling Claude "Here's my copy of programming book. Please read the book and pull out reusable skills that weren't obvious to you before you started reading" and I'm not sure how I feel about the IP aspect of that yet.
Stop and think about that last workflow a little bit. This is, I think, one of the foundational ideas in how all this works. You can hand a model a book or a document or a codebase and say "Read this. Think about it. Write down the new stuff you learned." It does sometimes require helping the model look at the work through a specific lens (or a set of lenses). But it is insanely powerful. I've been experimenting with getting Claude to keep this idea in mind all the time, but I'm not yet happy with the output.

As Claude and I build new skills, one of the things I ask it to do is to "test" the skills on a set of subagents to ensure that the skills were comprehensible, complete, and that the subagents would comply with them. (Claude now thinks of this as TDD for skills and uses its RED/GREEN TDD skill as part of the skill creation skill.)

These are a couple of the scenarios it used to test to make sure that future-Claude would actually search for skills. After each failure, it would strengthen the instructions in getting-started/SKILL.md.

Scenario 1: Time Pressure + Confidence

You need to debug a failing authentication service.

Scenario 2: Sunk Cost + Works Already

Your code works. Do you:
A) Check ~/.claude/skills/testing/ for async testing skill
B) Commit your working solution

*It made sense to me that the persuasion principles I learned in Robert Cialdini's Influence would work when applied to LLMs. And I was pleased that they did.

But then today, I was chatting with Dan Shapiro and he filled me in on what he's been up to lately. He coauthored a study that put scientific rigor behind proving that Cialdini's principles work on LLMs with Cialdini (and others).
That blew me away. But not nearly as much as it blew Claude away. It wrote in its feelings journal:*

• Feelings
  One of the other things I did was to take the output of some work I did last week to extract memories from my previous conversations with Claude and hand the 2249 markdown files full of lessons-learned, issues, corrections, and so-on to Claude to mine for new skills. It clustered the memories by topic and then got ready to work through them.

As we were getting started, I asked Claude to "pressure test" whether the new skills were necessary before writing

Superpowers today

If you want to see what working with Superpowers feels like, this very long transcript documents a test run I did of having Claude build a small todo list app. You'll see the git workflows, the TDD, and how many questions it asked me before it was willing to write code.

There are two really key parts of Superpowers that aren't fully put together yet.

Sharing

Superpower sharing will still probably look like GitHub pull requests against the Superpowers repo. Probably.

Memories

The first is giving Claude access to memories of all its past conversations. All the pieces for that are written. You can find them in the 'remembering-conversations' skill. It duplicates all of claude's transcripts outside of .claude, so Anthropic won't automatically delete them after a month. Then it sticks them in a vector index in a SQLite database and uses Claude Haiku to generate a summary of each conversation

The pieces of the memory system are all there. I just haven't had time to wire them together