Tag
#llm
The Skills I Actually Kept: My Claude Code Skills, and What I Use vs Build
The Claude Code skills I actually kept — article drafting, a full deploy pipeline, design-system contracts, self-testing — and the one rule I use to decide when to reuse a pre-built skill versus write my own: who's going to maintain it in six months.
UI Thrashing: The Reset Protocol That Actually Works
When AI-driven UI coding turns into an endless loop of micro-tweaks, don't negotiate with the model — kill the session, max the thinking effort, resend everything, and let it plan without more questions.
Claude Chat, Claude Co-Work, or Claude Code: How to Know Which One You Actually Need
The three tiers of working with AI — Claude chat, Claude co-work, and Claude Code — and the exact signal that tells you when it's time to move up a level.
One Button, No Subscription: How I Stopped Losing My Best Ideas
A physical single-button voice recorder turns my runs, my commute, and my meetings into first drafts — no subscription, and a human review still gates everything before it ships.
When GSD Became the Bottleneck, I Built Something Lighter
GSD asks a lot up front and forgets what it asked — great for serious multi-phase builds, terrible for rocket-ship mode. Here's the lightweight way of working I reach for instead: high effort, a two-item plan, round-based questions, big-bang execution, and two skills that auto-run the UAT.
MCP, Skills, Commands, Agents: A Product Manager's No-BS Guide to AI Buzzwords
A product manager's plain-language explanation of MCP, Skills, Commands, and Agents — revisited with the repo habits that came out of actually living with them: a docs folder as memory, granular Skills over plugins, CLI over MCP where it counts, and picking a model per task.
My AI Coding Stack: A Practical Guide to Using ChatGPT, Cursor, Claude Code, and Lovable
An evolving, practical AI coding stack — ChatGPT, Cursor, Claude Code, and Lovable — updated with how I now default to the Claude Code CLI, route Opus, Sonnet, and GPT to different jobs, and why I dropped Fable.
The Skill That Stops Claude From Forgetting Your Tasks
A new skill that turns a raw, unstructured brain-dump into a forced checklist of tasks for Claude — and makes it report back exactly what it did.
Please Stop Comparing LLMs With One-Shot Prompts
One-shot prompt comparisons between LLMs are everywhere right now, and they're useless. The problem isn't the model — it's that nobody bothered to write a spec.
Fable vs. Opus: My Token-Budget Verdict
I burned through two paid Max accounts in three days switching to Fable. Here's why I went back to Opus, and where Sonnet actually wins.
Write Your Acceptance Criteria Before You Let AI Touch the Code
A short AI-coding tip: write your acceptance criteria and tests during the plan, before any code gets written — so you never miss a requirement, and AI can test itself.
Let AI Test Its Own Work: A UAT Loop With Claude Code and Playwright
The practice I use to have Claude Code test its own work with a UAT file, Playwright, and per-persona .env credentials — then ask it to fix what it finds and retest, almost on its own.
One Database, One Repo: Why Multi-Repo Hurts AI Coding
A multi-repo split I made for scalability reasons turned into duplicated code and handover files. My rule now: one database, one repo, period.
Stop Prompting in English If It's Not Your Language
Prompting an LLM in English when it's not your native language adds a translation layer between your thinking and the model — switch to your own language and that layer disappears.
What If AI Sat in the Middle of Every Conversation?
A WhatsApp voice message and an old failed startup called Alfred got me thinking: what if AI sat between every sender and every receiver, translating each message into the style the other person actually wants?
Context Is Everything: Keeping AI Coding Agents From Losing the Plot
Why saturating a coding agent's context window quietly wrecks its output, and the docs/backlog/handover system I use to give agents real memory across a multi-repo project.
LLMs Don't Hallucinate — You Just Don't Ask Them to Show Their Work
The reframe that changed how I use AI for anything serious: it's not that the LLM hallucinates, it's that you never asked it to expand its context and prove its work. Here's why it happens — limited context, embeddings picking the closest match instead of the exact truth — and the two-step trick that kills most of it: expand the context, then demand proof for every sentence.
Permission Modes for AI Coding: Bypass, Auto-Edit, and Planning
Bypass, auto-edit, and planning are three different gears in Claude Code, not one setting — and knowing when to shift into planning is what saved a deploy bug I couldn't fix any other way.
Why AI Won't Save Us All (And Why That's Not What You Think)
AI hype hides the hard work of real products; designers, PMs, and engineers are still needed because making something great is the difficult 90 percent.
When ChatGPT Is Not Enough: How I Planned an XTerra with AI
Planning an XTerra triathlon revealed why ChatGPT struggles with long projects, and how moving to Claude Code with files and MCP fixed memory and dates.
AI Won't Replace You (Unless You Allow It To)
Treating AI as a collaborative companion rather than a one-shot replacement, and recognizing its limits to work faster and stay ahead.
Schools Are Missing the Biggest Opportunity AI Has Ever Given Us
A personal argument that schools should embrace AI to personalize learning for each child, using a five-step study framework as an example.