Forem: Sean K

6 Claude 4.5 Best Practices That Changed How I Work

Sean K — Sun, 04 Jan 2026 10:47:43 +0000

After months of daily Claude Code usage, here are the 6 practices that made the biggest difference.

1. Be Explicit

Don't: "Create an analytics dashboard"

Do: List exactly what you need with bullet points

Why it works: Claude fills in gaps with assumptions. More specifics = fewer surprises.

2. Give Context for Rules

Don't: "NEVER use ellipses"

Do: Explain WHY - "Your response will be read by text-to-speech, so never use ellipses since the engine won't know how to pronounce them."

Why it works: Claude follows rules better when it understands the reason.

Pro tip: Use CLAUDE.md for persistent context across sessions.

3. Long-Horizon Reasoning

Don't: Just say "Build this feature"

Do:

Use git commits for state
Write tests FIRST
Tell Claude NOT to wrap up early

Why it works: Complex tasks need checkpoints. Git commits let Claude recover context.

4. Avoid "AI Slop" in Frontend

Don't: "Build me a landing page"

Do: "No generic gradients, no excessive shadows. Reference: [screenshot]"

Why it works: Screenshots > descriptions. Show Claude what good design looks like.

5. Parallel Tool Calling

Don't: Let Claude run tools sequentially (8 seconds)

Do: Say "run these in parallel" (2 seconds - 4x faster!)

Why it works: Claude can execute multiple independent operations simultaneously.

6. Deep Research

Don't: "Research this topic" (lazy summarizer mode)

Do: "Search broadly, verify claims, triangulate sources. Don't stop at the first answer."

Why it works: Default behavior is to stop early. Explicit instructions unlock thorough research.

Bonus: Persistent Memory with Memento

Want Claude to remember these practices automatically? Check out Memento - it extracts session learnings into CLAUDE.md so you don't repeat yourself.

Follow me for more Claude Code tips: @_sean8

Memento: Give Claude Code Persistent Memory So You Stop Repeating Yourself

Sean K — Sat, 03 Jan 2026 20:46:50 +0000

If you use Claude Code, you've probably had this experience:

You: Install the dependencies
Claude: npm install
You: No, use pnpm in this project
Claude: pnpm install

Next session? Same thing. Claude has no memory between conversations.

The Problem

Claude Code reads CLAUDE.md files at the start of each session - that's how it learns your project conventions. But manually maintaining these files is tedious, and you often forget to document the corrections you make during coding sessions.

The Solution: Memento

I built Memento - a Claude Code command that analyzes your conversations and extracts actionable insights automatically.

Named after the Christopher Nolan film where the protagonist leaves notes for his future self (because he can't form new memories), Memento helps you leave notes for future Claude sessions.

How It Works

At the end of any coding session, run:

/memento

Memento will:

Analyze your conversation for corrections, preferences, and learnings
Filter for actionable insights (not vague observations)
Categorize into project-specific vs. personal preferences
Present suggestions for you to approve

You select which ones to keep, and they're appended to the appropriate CLAUDE.md file.

What Makes a Good Suggestion?

Memento focuses on actionable insights:

✅ Actionable	❌ Not Actionable
"Use pnpm, not npm"	"Check package manager"
"Tests are in `__tests__/`"	"This was a good session"
"Always show command before running"	"Be more careful"

Installation

One-liner:

mkdir -p ~/.claude/commands && curl -fsSL https://raw.githubusercontent.com/SeanZoR/claude-memento/main/.claude/commands/memento.md -o ~/.claude/commands/memento.md

That's it! The command is now available in all your Claude Code sessions.

Try It Out

Give it a spin and let me know what you think! Issues and PRs welcome.

SeanZoR / claude-memento

Memory tools for Claude Code: /memento extracts session learnings, /brag generates social content ideas

███╗   ███╗███████╗███╗   ███╗███████╗███╗   ██╗████████╗ ██████╗
████╗ ████║██╔════╝████╗ ████║██╔════╝████╗  ██║╚══██╔══╝██╔═══██╗
██╔████╔██║█████╗  ██╔████╔██║█████╗  ██╔██╗ ██║   ██║   ██║   ██║
██║╚██╔╝██║██╔══╝  ██║╚██╔╝██║██╔══╝  ██║╚██╗██║   ██║   ██║   ██║
██║ ╚═╝ ██║███████╗██║ ╚═╝ ██║███████╗██║ ╚████║   ██║   ╚██████╔╝
╚═╝     ╚═╝╚══════╝╚═╝     ╚═╝╚══════╝╚═╝  ╚═══╝   ╚═╝    ╚═════╝

Memory tools for Claude Code
Because Claude forgets, but your notes don't.

Install

mkdir -p ~/.claude/commands && \
curl -fsSL https://raw.githubusercontent.com/SeanZoR/claude-memento/main/.claude/commands/memento.md -o ~/.claude/commands/memento.md && \
curl -fsSL https://raw.githubusercontent.com/SeanZoR/claude-memento/main/.claude/commands/brag.md -o ~/.claude/commands/brag.md

Commands

`/memento` — Extract session memories

Run at the end of any Claude Code session to capture learnings into your CLAUDE.md files.

/memento [--project PATH] [--user PATH]

Argument	Description
`--project PATH`	Custom project memory location (default: `./CLAUDE.md`)
`--user PATH`	Custom user memory location (default: `~/.claude/CLAUDE.md`)

What it does:

Analyzes your conversation for mistakes, corrections, and preferences
Filters for actionable insights only
Sorts into Project Memory (./CLAUDE.md) vs User Memory (~/.claude/CLAUDE.md)
Lets you pick which suggestions…

View on GitHub

Have you built any tools to improve your Claude Code workflow? I'd love to hear about them in the comments!

I Accidentally Made Claude Ask Itself the Same Question 1,966 Times

Sean K — Sat, 03 Jan 2026 20:26:03 +0000

I Accidentally Made Claude Ask Itself the Same Question 1,966 Times

Last night I broke Claude Code in the most absurd way possible.

The Setup

I've been experimenting with Claude Code, Anthropic's CLI tool for AI-assisted development. One of its features is plugins, and I installed one called Ralph Wiggum - a loop mechanism that lets Claude iteratively work on tasks.

The idea is simple: give Claude a task, let it loop until it's done. Great for complex, multi-step work.

The Prompt

I wanted Claude to help me build out my personal brand. So I typed:

Create an amazing full stack of branding presence for me

And enabled the Ralph Wiggum loop.

What Went Wrong

Here's the loop configuration file that got created (.claude/ralph-loop.local.md):

---
active: true
iteration: 1966
max_iterations: 0
completion_promise: null
started_at: "2026-01-03T16:21:24Z"
---

Create an amazing full stack of branding presence for me

See the problem?

max_iterations: 0 = unlimited iterations

The Symptom

Every single message I sent got hijacked. Claude would respond, but then immediately get interrupted with:

Stop:Callback hook blocking error from command: "callback":
Create an amazing full stack of branding presence for me

Over. And over. And over.

1,966 times.

I couldn't have a normal conversation. Every response was intercepted. Canceling the loop via the plugin command failed with a parse error. I was trapped in branding purgatory.

The Irony

The task was to create my branding presence.

Claude attempted this task 1,966 times without producing a single asset.

This might be the most committed non-delivery in AI history. Maximum effort, zero output.

The Fix

Eventually Claude itself helped me debug. We:

Searched for hook configurations
Found the .claude/ralph-loop.local.md file
Discovered it had been running for hours
Deleted it with rm

rm .claude/ralph-loop.local.md

Instant relief.

Lessons Learned

1. Always Set Max Iterations

If you're using any loop mechanism, always set a sensible limit. max_iterations: 0 sounds like "none" but means "infinite."

2. Know Where Config Files Live

Claude Code stores loop state in .claude/ directories. When things go wrong, check there first:

~/.claude/settings.json - global settings
.claude/settings.local.json - project settings
.claude/ralph-loop.local.md - loop state (if using that plugin)

3. Hooks Can Hijack Everything

Claude Code hooks are powerful - they can intercept and modify any interaction. A misbehaving hook can make the entire tool unusable.

4. Vague Prompts + Infinite Loops = Bad Time

"Create an amazing full stack of branding presence" is vague. Claude kept asking for clarification, which triggered another loop iteration, which asked for clarification...

The Silver Lining

I didn't get my branding assets. But I got a great story, a deeper understanding of Claude Code internals, and content for dev.to and Twitter.

Sometimes the bug is the feature.

Have you broken an AI tool in a spectacular way? Drop your stories in the comments.

P.S. - I still need that branding presence. Maybe I'll try again... with max_iterations: 5.

Getting a NO from an open source maintainer

Sean K — Fri, 05 Dec 2025 12:35:47 +0000

You think you're immune from rejections in reallife, zen and all, but getting a big NO from an opensource maintainer always hits differently

Forem: Sean K

6 Claude 4.5 Best Practices That Changed How I Work

1. Be Explicit

2. Give Context for Rules

3. Long-Horizon Reasoning

4. Avoid "AI Slop" in Frontend

5. Parallel Tool Calling

6. Deep Research

Bonus: Persistent Memory with Memento

Memento: Give Claude Code Persistent Memory So You Stop Repeating Yourself

The Problem

The Solution: Memento

How It Works

What Makes a Good Suggestion?

Installation

Try It Out

SeanZoR / claude-memento

Memory tools for Claude Code: /memento extracts session learnings, /brag generates social content ideas

Install

Commands

/memento — Extract session memories

I Accidentally Made Claude Ask Itself the Same Question 1,966 Times

I Accidentally Made Claude Ask Itself the Same Question 1,966 Times

The Setup

The Prompt

What Went Wrong

The Symptom

The Irony

The Fix

Lessons Learned

1. Always Set Max Iterations

2. Know Where Config Files Live

3. Hooks Can Hijack Everything

4. Vague Prompts + Infinite Loops = Bad Time

The Silver Lining

Getting a NO from an open source maintainer

`/memento` — Extract session memories