Mark Phelps

Got tired of this email from GitHub, and of slow, expensive macOS CI runners. Then I remembered I have a capable Mac Mini at home. So I made a skill that sets up self-hosted runners on Linux or Mac. Give it to your agent and save some money: https://github.com/markphelps/skills

17 Apr

Before AI: I moved slow, but I knew every line. Now: I move fast, and I'm auditing things I didn't write. Both ways are exhausting.

13 Apr

You can just do things. Built a self-hosted cora.computer for email: - auto-triages into custom labels w/ AI - marks processed as read, uncertain → Needs Review - queues archive proposals for approval Runs on Cloudflare Workers + D1 + AI Gateway (Gemma 4)

13 Apr

Also added: - signed CLI for ops (summary, approve/reject, feedback, retention) - feedback memory so corrections sharpen future labeling - Skill file for Hermes to ping me twice daily reports

6 Apr

Use Google's latest OSS model Gemma-4 in Hermes or OpenClaw via @cloudflare workers AI It's seriously impressive and noticeably faster than Kimi in my testing

7 Mar

I got tired of reviewing PRs blind. So I built Sherpa - a Chrome extension that explains GitHub PRs as you read them. On demand. Your code never leaves your machine. New post on how it works: https://markphelps.me/posts/building-llm-powered-code-review-sidebar/

25 Feb

deepwiki is so good ask questions about a codebase in natural language no clone, no boot up, fast response with sources 🧑‍🍳😘

24 Feb

Apple finally approved my FamilyControl prod permissions. Going to finally launch Stay on TestFlight this week. Related, I migrated trystay.app to @Cloudflare with OpenNext in 5 mins from the Triangle Co. Saves $40/month.

21 Feb

Look how cute this lil dude is. This is much easier than building mobile apps ngl

21 Feb

Me and Claude cooking this morning.. Who knew building Chrome extensions was this fun

20 Feb

Gave Sherpa a fresh coat of paint and now it's running on @Cloudflare Workers + Durable Objects so summaries stay synced across tabs and users in real-time

20 Feb

vibe coded a chrome extension that explains PR changes as you scroll through them near-realtime what a time to be alive

20 Feb

Adding devcontainer support to Flipt let me use --dangerously-skip-permissions with Claude Code. Now the agent can run wild inside the container. Can't touch my host. Mounted my skills in. Full control. Wrote about why this matters: https://markphelps.me/posts/running-ai-agents-in-devcontainers/

19 Feb

Show me your OpenClaw agents. Here's mine: Hunter - Blog & social posts Gilfoyle - Server maintenance 🤘 George - Flipt business. British, organised David G - Workout planning & accountability Dwight - Assistant to the regional manager. Coordinates them all

19 Feb

cc @georgemac.bsky.social

14 Feb

Codex 5.3 is scary good.. and fast. Vibed this landing page using some frontend skills from skills.sh Deployed to Cloudflare Workers using the Cloudflare Docs MCP server https://openvault.agentty.sh/ We're cooked gang

9 Feb

OpenClaw Mission Control update: usage graphs 1.1M tokens across 4 agents in 6 days main: ~500k 🐷 coding: ~200k social: ~30k blog: ~15k Models: opus-4-6: 45% gpt-5.2-codex: 35% MiniMax: 15% Now I know exactly where my API budget goes.

Everyone's building OpenClaw mission controls. I built one too, but with a twist: It shows you what OpenClaw hides—the raw LLM inputs, tool calls, token counts, everything. A thread 🧵

Built with: 🛠️ NextJS 🗄️ Supabase 🎨 Shadcn 🤖 Claude Code Opus 4.6 Shipped faster than I expected. The scaffolding does most of the work.

The best part: seeing what the LLMs actually return. - Tool calls with full payloads. - Structured outputs. - Token breakdowns per turn. - Success/fail metrics. You spend all day talking to agents. Never see what's under the hood. Now you can.

Everyone's building OpenClaw mission controls. I built one too, but with a twist: It shows you what OpenClaw hides—the raw LLM inputs, tool calls, token counts, everything. A thread 🧵

OpenClaw is great, but it's a black box. You see the response. You don't see: - What the LLM actually processed - How many tokens per turn - Which tools fired - Success/failure rates I wanted visibility. So I built it.

Mission Control: - Activity timeline - every message, tool call, cron job - Calendar - 12 scheduled jobs, all tracked - Full-text search across ALL conversations + transcripts - Raw data panel - see exactly what the LLM saw

2 Feb

I've installed and nuked clawdbot/openclaw like 4 times now. Local linux, then VPS. Each time something broke badly enough that nuking and starting over was the only option.

2 Feb

First the clawdbot → openclaw transition Then trying to get multiple agents talking to each other People are selling openclaw setups to non-tech users as a service It ain't ready for that...

2 Feb

If I'm struggling this much to get a setup I love, non-techies are gonna have a bad time Or could just be a skill issue

1 Feb

I'm building an MCP gateway in Rust 🦀 Instead of loading every tool upfront, your agent gets four meta-tools: search, inspect, call, list. Progressive discovery. 50 tokens vs 18,000 for typical workflows. Useful or solving a problem that doesn't exist?

30 Jan

i think our jobs are safe for a little while longer..

25 Jan

Spent a few hours yesterday bringing my old blog back to life: https://markphelps.me Go/Hugo -> Astro AWS S3 -> Cloudflare R2 Netlify -> Cloudflare Pages Thumbnails conjured via Replicate API I have been fully absorbed into the orange cloud

25 Jan

Now I just gotta find the time to start writing new posts...

9 Jan

Opus is a beast at planning (and coding) but I still have it spawn parallel subagents using Sonnet 4.5 when possible because Sonnet also ships code like a machine. Save on those rate limits!

7 Jan

@Experian is impossible to reach when your account gets hacked. Their AI blocks you from any human help. Anyone dealt with this? WDYD?

6 Jan

I don't and well never understand Nix and I dont care who knows it!

4 Jan

thinking about building a simple app that tracks car maintenance + deadlines (registration, inspection, insurance renewals). i keep a spreadsheet (like a nerd) but never look at it until something's already overdue. anyone actually solved this or is everyone just vibing?

4 Jan

One thing I've noticed over this holiday break: I've been rapidly prototyping ideas with AI, then taking a full day to sit with them and decide if they're actually worth pursuing. And since the code is AI-generated, killing bad ideas feels completely painless.

2 Jan

Are we already at the point where foundational models have become commodities?

2 Jan

I want this page as an API https://yourgpt.ai/tools/openai-and-other-llm-api-pricing-calculator Does this exist? Up to date LLM model prices/context limits/etc info as JSON

2 Jan

Man i do miss those 2x usage limits today.. come on Anthropic please extend until most people go back to work on Jan 5

Here's everything I've learned while 'vibe coding' react native iOS apps this year (most of this applies to any code, not just RN/mobile):

Finally, ABC - Always Be Clearing context. Compaction is the devil. If you've documented your plans and have a task tracker like Beads, you can nuke context whenever you want - this is the secret to cleaner code and sharper design.

What are you building with AI in 2026? Drop your best tips below 👇

Here's everything I've learned while 'vibe coding' react native iOS apps this year (most of this applies to any code, not just RN/mobile):

create subplans for large feature work before touching any code. this one for Claude Code is a game-changer: https://github.com/obra/superpowers Add a task tracking system like https://github.com/steveyegge/beads to track bugs, future enhancements, etc. You can now ship like a team of 5+

models are getting so good that less is more. you don't need a 2,000 line CLAUDE.md - quality drops after a certain point due to context limits. every so often ask Claude to trim its own CLAUDE.md file

Here's everything I've learned while 'vibe coding' react native iOS apps this year (most of this applies to any code, not just RN/mobile):

always start with a PRD and reference it in your CLAUDE.md/AGENTS.md the PRD is the source of truth when everything gets messy update it continuously as you learn what works Claude (web) is surprisingly good at creating these with the right prompt

use mobbin or dribbble to steal visual inspiration. drop screenshots in a 'design/influence' folder. then let your agent extract a design system - colors, icons, spacing, animations, the works. this lives right next to your PRD as the visual source of truth

Here's everything I've learned while 'vibe coding' react native iOS apps this year (most of this applies to any code, not just RN/mobile):

don't abstract too early-resist extracting components from a single screen until you're certain they'll be reused (atoms like buttons are the exception)

25 Dec 2025

calling it now: 2026 is the year of brutalist apps

23 Dec 2025

Gemini 2.5 Flash is absurdly fast and honestly holds its own against the bigger Gemini models.. at least for UI/design work. What are your thoughts chat?

22 Dec 2025

ask your ai agent to recommend 2-3 pre-commit hooks for your current project this + husky = vibe coding bliss

22 Dec 2025

CodexBar by @steipete is super nice. - works with Codex, Claude Code and Gemini - open source: https://github.com/steipete/CodexBar this feels like the future, AI helping us build polished little apps that do one thing really well.