Take The Claude Pill @taketheclaudepill - Tumblr Blog

Context is full

I was chatting with a dev friend about AI and I asked how he felt. "Obsolete, but empowered," was the response.

I'm struggling with how infrequently I open my true IDE these days, opting for zed because it is lightweight and handle *.md pretty well. I'm struggling with what it means to be a developer that doesn't really need to write code because describing my architecture is enough. I'm struggling with context switching because now I have time to work on three things at once because I'm not actually working on three things at once - until all at once I'm working on three things at once.

Orchestration Practice

I spent a couple hours (and $2.27) building a command line Orchestrator that does the following: * Retrieves a list of open issues that are tagged agent-ready. * Determines the best one to work on (using Haiku) * Creates a plan (using Sonnet) * Implements the plan (again using Sonnet) * Opens the PR

I opted to use github issues rather than Asana to make it easier to test with (2 API instead of 3). I chose to work directly with the Claude API for the same reason...speed to ship.

At this point it has handled 1 single issue. I had about a half dozen false starts (probably more...)

My go forward plan is to stand up repos for most projects specifically for AI use as well as preserve my human versions. Then, I'd like to let it run nightly on each repo. I'll have to update to pull in issues from Asana, and providing suggested files to work on makes a HUGE difference. But, for minor bugs I feel good trying to let it handle overnight. It does have execution limits to prevent loop run away as well as some really hard rails on what it can do.

An experiment with local models when Claude put me in timeout

Local models suck. That is probably not entirely true...but wow I've never heard my computer fans work like that. I toyed around with a local version of qwen3 yesterday as my model in Claude Code. It took ~3 minutes to start doing anything after my initial ask (refactor a specific method in lrpbot repo). I'm sure that is because the context was enormous. But, that kinda makes it a non-starter. I think I might have a better experience using a local model in my IDE rather than Claude Code, but honestly I don't think I've ever hit a limit with copilot - so why bother?

My intent was a local functional-ish option for the times I run out of allocation on Claude. I think more useful would be a way to fall back to OpenAI. Why not use just use codex? Well, codex is fine but I have a fair bit of tooling around and in Claude Code that I miss in codex.

This article suggests 8 levels of agentic engineering. However, it lumps 1 & 2 together (tab complete and agentic IDE)...so that leaves 7 and then it says no one is really doing 8, okay fine...down to 6. But, I am struggling with nuance around 3-7. Specifically 3-5 feel like one level and 6-7 feel like another. All that being said, I do see the different layers. I would re-label as 1. AI utilization in creation of code. (this is 1 & 2 from the article). 2. Integrating systems (MCP, skills, context .md files) and "co-working" with AI. Not just directing what code should be written, but using AI to help refine the direction before code is written AS WELL as respond when it's written. 3. Entrusting agents with experimentation and, on success, moving forward with options that I have given little or no thought. So...I think I'm successful at #1 and have been for awhile. I'm waist deep in #2 and have added practices in the last month that are improving this. To move forward on #3 requires going beyond a tool like Claude or Copilot and into API space within a repo.

Much to think about.

AI's coding ability is outpacing our ability to wield it effectively. That gap closes in levels — 8 of them. Here's the progression from tab

Chat vs. Cowork vs. Code

Spotted this graphic on an X post Found it very helpful in distinguishing between the 3 lanes of Claude. I've been freely using all of those lanes. Pretty haphazardly. And if I'm being honest about it, it had created a bit of a sprawling chaotic situation in my Claude setup. I'm not remembering where I installed certain skills. It's basically however I'm using Claude at that moment. Then a week later, I don't remember where that skill is installed or even that I installed the skill. This feels reminiscent of early WordPress builds where I would end up with 3 plugins that did social media sharing. Other challenges are I'm not sure what the rules are regarding where a skill is available to you when you install it. If I install it via Claude chat, is it also available in cowork? Similar thing with prompts. Did I start the prompt in code or in a chat project? There's a lot of talk about organizational governance of A.I.. But I think I'm in need of some individual usage governance.

In this video, I walk through my experience using AI to resolve npm audit vulnerabilities in a project, where I started with 37 vulnerabilit

In this video, I walk through my experience using AI to resolve npm audit vulnerabilities in a project, where I started with 37 vulnerabilities and ended up reducing them significantly. I demonstrate how running 'npm audit fix' multiple times toggled between 11 and 53 vulnerabilities, prompting me to leverage cloud code for a more efficient solution. By allowing the AI to determine the necessary package updates, I was able to resolve the issues in about six minutes, which would have taken me much longer manually.

At the end I show a PR opened by Claude Code using the existing pull_request_template from the repo.

Using AI to jumpstart the LRP X account

Over the weekend, I decided to see how A.I. could help me get the Lone Rock Point X account going. That's an account that I haven't done anything with over the years. There really is limited strategic value and I'm not sure there really is now, but I thought it would make for an interesting use case in seeing how A.I. can help deliver in an area that has long been backburnered, even ignored. Here's how I approached it. I used a Claude skill call "copywriter" (not sure where I got that skill, but it's in my Claude configuration) and asked it to help me create a 30 X posts relevant to what LRP does. It proceeded to ask me a few question then generated me a list of 20 draft posts. I reviewed that list. Indicated about 6 I hated, 5 that I loved and everything was average. Then based on that feedback, try again. Received a revised list of 20 posts. That looked good so I told it to sprinkle those posts across a posting schedule for the rest of March. Schedule 2 or 3 posts per weekday and 1 on each weekend. After that, I looked at we have and told it we don't take ourselves so seriously all the time, add in 20 "fun posts", then it went through the process of finding that right "fun tone" and drafting a list. After some review and revision, I merged the list with the "serious list" and now we have a posting schedule for about 40 days. One serious post and one fun post each day. It suggested hashtags and even made image suggestions. I then told Claude to convert that process into a skill that I could invoke the next time I wanted to create 30 more days of scheduled X content. I took the image suggestions and starting seeing if we could auto generate those. Claude is terrible at creating images. In fact right it doesn't seem like it can. What it is really good at is creating prompt templates that you can paste anywhere. You can provide it 3-5 sample images you like or that have a style you want to emulate and ask it to create a reusable prompt template which you can paste into Google Gemini (which is available with our LRP Google Workspace account), Canva, and ChatGPT. I'm sure you could paste them into other solutions. I don't know what Figma has in this area. So depending on the model, the results differ. Canva is terrible at generating anything with text. Gemini thrives there. Generating images is a little too manual right now. It would nice to be able to bulk create a set of images and assocaite them with scheduled X posts. Right now, I'm generating them, then uploading them to a drive folder and updated the Google Sheet of scheduled X posts to know what image file goes with the scheduled post. With all that information defined, I can automate the actual posting on a daily schedule. All the pieces are in places for that to run on autopilot - Post text, hashtags, image.

Include scheduled tasks in the Claude ecosystem

Apparently, Claude has the ability define and schedule repeating tasks. This is like it having its own cron system. Good breakdown here: https://artificialcorner.com/p/claude-automates?hide_intro_popup=true

Naturally, there's multiple ways to work with them.

Via the Claude Chrome Extension

In Claude Cowork as part of Claude Desktop.

I wonder if that means Claude desktop needs to be running on your machine to execute on schedule.

A couple years ago, I was at an industry event of CMS vendors. Naturally, A.I. was the topic du jour and at the time prompting was everything. Someone joked that someone needs to implement a CMS to be a knowledge library of good prompts. It was a joke, but I thought that was good idea.

Fast forward a couple years. We've moved beyond prompts. I personally have skills, plugins, and subagents installed and spread across my own Claude implementation of Claude AI, cowork, and code. It feels like genuine sprawl. I can't imagine trying to scale out sharing of best practices to a team. How do we inventory our shared skills and subagents? Maybe it really is time for a CMS to be a knowledge base for this. Think about it, CPTs in WordPress to catalog skills, plugins, scheduled tasks, etc. For larger organization, I can see a rating/star system having some value add as well.

#scheduled tasks

Trending Blogs

Recently Viewed Blogs

Take The Claude Pill