✦

Everyone's Saving Prompts. Nobody's Building Systems.

3 prompts that actually persist + the system that runs them without you.

You commented SYSTEM. Here's everything.

There are 3 Claude Code mega-prompts going viral right now. They cover self-learning research, multimodal memory, and persistent context. The ideas are sound. The problem: you paste them, Claude runs them, the session ends, and tomorrow Claude has no memory any of it happened.

Below are those 3 concepts rebuilt as proper skills that persist across sessions, plus the 4-layer system that makes them run automatically without pasting anything.

SET THIS UP FIRST

The 4-Layer System

The prompts below are powerful. But they only work once without this foundation. Set up these 4 layers and every prompt you run from this point persists automatically.

system-stack.sh

        Layer 1: CLAUDE.md

          Your brand, voice, rules. Loaded every session.

        Layer 2: Project folders

          Switch dirs = switch context. No re-explaining.

        Layer 3: Skills

          Capabilities that auto-activate when relevant.

        Layer 4: Domain rules (CARL)

          Right rules load for the right task.

        // Built once. Runs every session.

CLAUDE.md docs: docs.anthropic.com/en/docs/claude-code/memory

CARL (domain rules): github.com/ChristopherKahler/carl

PROMPT 01

Build a Research Radar

Claude monitors your niche and only flags what's actually new.

The viral version of this prompt tells Claude to "find new information." That's too broad. Claude wastes tokens pulling things you already know. This version filters against your existing knowledge base first, so you only see what's genuinely new or contradicts something you believed.

research-radar.md

        Create a skill at .claude/skills/research-radar/ that does one thing: find information I don't already have.

        Before searching, read my CLAUDE.md and any existing notes in /research/known-landscape.md to understand what I already know about my niche.

        Then search web, Reddit, Hacker News, and relevant forums for: new tool launches, strategy shifts, competitor moves, and workflow changes in [YOUR NICHE].

        For every finding, check it against known-landscape.md. If it's redundant, discard it. If it contradicts something I believe, flag it as a challenge. If it's genuinely new, add it.

        Store validated findings in /research/radar-[date].md with: source URL, one-line summary, and whether it's NEW or CHALLENGES existing knowledge.

        End each run by updating known-landscape.md with confirmed new patterns. Old findings get promoted. Noise gets deleted.

Why this is better: The viral version dumps everything it finds into one file. This version filters against what you already know, so you only review what matters. It also flags contradictions, which is where the real insights are.

PROMPT 02

Build a Work Log That Searches Itself

Every file you create gets tagged, described, and made searchable by meaning.

The viral version uses vector embeddings and ChromaDB. That's powerful but overkill for most people. This version creates a structured log that Claude can search natively without any external dependencies. No database setup. No API keys. Just files that describe themselves.

work-log.md

        Create a skill at .claude/skills/work-log/ that automatically catalogs everything we create or reference in a session.

        After every session, generate an entry in /work-log/entries/[date].md containing: files created or modified (with paths), a 2-sentence description of what each file does, tags for content type (doc, code, image, data, config), and the task context (what we were trying to accomplish).

        Maintain a master index at /work-log/index.md that lists every entry with tags and descriptions. When I ask "find the thing we built for X" or "where's that file about Y," search the index by description and tags first, then scan entries if needed.

        Update my CLAUDE.md to include: "At session start, read /work-log/index.md. Reference past work when relevant to current tasks. At session end, log what was created."

Why this is better: No external dependencies. No database. No API keys. Works with Claude Code out of the box. And by logging task context (not just filenames), Claude can find things by what you were doing, not just where you put them.

PROMPT 03

Build a Memory That Promotes Itself

Claude distills what it learned today into permanent knowledge tomorrow.

The viral version creates three memory files and a nightly cron. The idea is right, but the execution dumps everything into long-term memory with no filter. After a month, your memory file is bloated with noise. This version uses a two-stage promotion system: observations stay temporary until they repeat. Only patterns get promoted.

memory-promoter.md

        Create a skill at .claude/skills/memory-promoter/ with two memory layers:

        Layer 1: /memory/observations.md — temporary. At the end of every session, extract: decisions made, preferences expressed, corrections given, patterns noticed. Each entry gets a date and a tally counter starting at 1.

        Layer 2: /memory/permanent.md — promoted. Run a weekly review that scans observations.md. Any observation that appears 3+ times across different sessions gets promoted to permanent.md with a one-line summary of the pattern. Observations that haven't repeated in 30 days get archived to /memory/archive.md.

        Update CLAUDE.md to: "At session start, read /memory/permanent.md in full. Scan /memory/observations.md headers only. Apply permanent patterns to all output. Flag when current work contradicts a permanent pattern."

        The goal: Claude remembers what matters and forgets what doesn't. No bloat. No noise. Only patterns that proved themselves.

Why this is better: The viral version promotes everything. After a month, your memory file is 500 lines of noise. This version only promotes patterns that repeat 3+ times. Observations that don't repeat get archived automatically. Your permanent memory stays lean and accurate.

The Key Difference

The viral prompts are one-shot. You paste them, Claude builds the thing, the session ends. Tomorrow Claude doesn't know any of it exists.

The prompts above are different. They create skills (files in .claude/skills/) that persist on your machine. They update your CLAUDE.md so Claude reads them automatically. They reference each other. The research radar feeds the work log. The work log feeds the memory promoter. The memory promoter updates CLAUDE.md. The cycle runs itself.

Prompt approach: Paste → Run → Session ends → Gone
System approach: Build once → Runs every session → Compounds over time

A prompt is a conversation. A system is a machine. Conversations end. Machines run.

The Order

1

Write your CLAUDE.md

Brand, voice, audience, rules. 15 minutes. This is the foundation every skill reads.

2

Run Prompt 3 first (Memory Promoter)

This creates the memory layer that everything else feeds into. Without it, nothing persists.

3

Run Prompt 2 (Work Log)

Now everything you create gets cataloged and searchable. Claude stops losing your work.

4

Run Prompt 1 (Research Radar)

Last because it needs the other two to store and remember its findings. Now Claude learns on its own.

Run all 3 prompts once. They create skills that live on your machine. Tomorrow when you open Claude Code, all three are already loaded. No pasting. No repeating. The system runs.

Coming next

My actual CLAUDE.md file — the one that powers everything above

Every rule. Every integration point.
Copy it and make it yours.

Follow @mikemeansbusiness_ai

Reply with which prompt you're running first. I'll tell you what to put in your CLAUDE.md to make it work best.

— Mike

✦ Mike Means Business · @mikemeansbusiness_ai

Build systems rather prompts