How to Build a Self-Improving AI Operating System With Claude PRISM

I run a content agency. We produce articles, Knowledge Panel sprints, social campaigns, and influence audits for founder-led businesses. Before PRISM, every Claude Code session started from zero. I’d explain who I am, what my agency does, who my clients are, and how we work. Every. Single. Time.

That repetition costs hours per week. Worse, the knowledge never compounds. Claude would discover something useful in one session, like a better article QA process or a smarter way to handle client onboarding, and then forget it by the next conversation. Every insight died at the end of the session.

PRISM fixes that. It gives Claude Code persistent memory, structured SOPs that update themselves, automatically referenced methodology frameworks, and an identity layer that tells it who you are from the first message. The system gets measurably better every time it runs.

This guide walks you through setting it up, what each component does, and how to get the most out of it.

Install the PRISM in Under Five Minutes

Three commands. That’s all it takes.

git clone https://github.com/Jwendt0324/PRISM.git ~/Documents/Claude/PRISM
cd ~/Documents/Claude/PRISM
./setup.sh

The setup script creates your personal identity folder, generates a context file from the template, prompts for your name, email, and company name to personalize the automation scripts, and wires PRISM into Claude Code so it loads automatically on boot.

After setup, open Claude Code anywhere on your machine and say “Run the identity scan.” Claude will scan your local files, Gmail, Google Drive, calendar, and GitHub to build a comprehensive picture of who you are, what you work on, and who you work with. You choose which sources to include. Everything stays local in a .personal/ folder that never gets committed to git.

That scan is what transforms Claude from a generic coding assistant into something that knows your business.

Understand the Two-Layer Architecture

The PRISM uses two layers that work together but stay separate.

The shared layer is the repo itself. Skills, SOPs, canonical frameworks, hooks, scripts, and templates. This is version-controlled and synced across your team. Everyone gets the same methodology, the same quality standards, the same automation.

The personal layer lives in .personal/ and is gitignored. Your identity context, your memory bank, your scan preferences. When you clone the repo and run setup, you get your own personal layer. Your teammate gets theirs. The methodology is shared. The identity is private.

Claude Code loads your personal CONTEXT.md at the start of every session. That file tells Claude your name, role, clients, priorities, communication style, and what you’re working on right now. From there, Claude automatically routes tasks to the appropriate skill or SOP.

No more explaining yourself every session.

Learn What Each Component Does

Skills

Skills are action-oriented prompts that Claude loads and follows when it recognizes a matching task. There are 15 of them. Article writing, content repurposing, Knowledge Panel sprints, Dollar-a-Day campaign management, prospect follow-up, influence audits, and more.

Each skill has YAML frontmatter, trigger keywords, step-by-step execution logic, quality rules, and cross-references to the canonical frameworks it draws from. When you tell Claude to write an article, it doesn’t guess how. It loads the article writer skill, checks for approved examples to match quality against, verifies the content tree for overlap, writes the piece, and runs it through an 18-step quality gate.

You can add your own skills by dropping markdown files into the skills/ directory.

SOPs

SOPs are the deeper reference documents behind the skills. Thirty-four of them cover client work, business operations, and file management. They include qualifying checklists, execution steps, verification procedures, human gates, and anti-vandalism safeguards.

The key difference between a skill and an SOP: a skill is what Claude loads and executes. An SOP is the full reference it consults when the skill doesn’t cover an edge case. Skills are condensed. SOPs are comprehensive.

SOPs are living documents. When Claude discovers a better approach or a new pitfall during a session, it automatically updates the relevant SOP. Version numbers bump. Dates update. The system gets smarter.

Canon Frameworks

The blitzmetrics-canon/ directory holds 26 canonical methodology files. These are the source of truth for everything the PRISM does. Content Factory (the six-stage process from Plumbing through Perform), Dollar-a-Day advertising strategy, the Topic Wheel for content organization, the Nine Triangles framework, MAA reporting, GCT discovery, and more.

If an SOP ever contradicts a canon file, the canon wins. That hierarchy prevents methodology drift as the system grows.

Hooks

Thirteen automation hooks run in the background during every Claude Code session. They don’t require any action from you. They just work.

Pre-tool guard blocks destructive commands before they execute. Try to run rm -rf ~/ or force push to a protected branch, and the hook catches it with a warning before any damage happens.

Session start injects daily context, creates log directories, and primes Claude with statistics about recent activity.

The post-tool logger records every tool call in a daily action log. This creates an audit trail you can review, analyze, or feed into reports.

Post-tool backup automatically creates .bak copies of files before they’re overwritten. You can always recover the previous version.

Prompt submits filters for confidentiality topics and suggests relevant skills based on what you’re asking about.

Session end parses the full session transcript into both machine-readable JSONL and human-readable markdown summaries.

There are seven more covering compaction preservation, task completion logging, stop summaries, failure recovery, Telegram notifications, and file naming enforcement. All of them run silently unless they need your attention.

Scripts

Twelve utility scripts handle operational tasks. Generate end-of-day reports with a single command. Run system health checks that audit your hooks, logs, rules, and good-examples library. Map your content tree to detect overlap and cannibalization. Export conversations to shareable formats. Extract session metrics for cost and usage tracking. Rotate old logs automatically.

These scripts are the operational backbone. They turn raw session data into actionable intelligence.

Team Operating System

Nineteen template files cover everything you need to run a team: directory, communication protocols, decision rights matrix, meeting cadence, delegation framework, onboarding kit, role scorecards, tool stack documentation, project management, client communication, escalation playbook, and scaling plan.

These ship as example files from a real agency deployment. Each one has an EXAMPLE banner at the top. Replace the placeholder data with your team’s information, delete the files you don’t need, and you have a complete team operating system that Claude can reference when handling team-related tasks.

Overnight Automation

Six overnight prompts let you kick off deep analysis jobs before bed and review the results in the morning. Memory bank refresh, team operating system audit, content audit, full system overhaul, methodology alignment check, and system upgrade planning.

These are the compound interests of the PRISM. Run them weekly or monthly, and the system self-improves without requiring your active attention.

Run Your First Real Task

Once setup is complete and the identity scan has run, try this:

Open Claude Code in any directory and say: “Write an article from this transcript.” Paste or link a video transcript. Claude will load the article writer skill, check for existing content on the topic (anti-vandalism), write the article in line with your voice profile, run the 18-step quality gate, flag any human-review items, and output a formatted document ready for WordPress.

Or try: “Run a health check on my PRISM.” Claude will audit your hooks, logs, rules, stale files, and disk usage, then generate a report with specific recommendations.

Or: “Create a new SOP for [process you repeated today].” Claude will pull the SOP creation template, walk through the structure, and file the finished document in the right directory with the index updated.

Every task feeds back into the system. Session logs capture what was done. Those logs inform SOP improvements. Those improvements make the next session faster and more accurate.

Customize It for Your Business

The CUSTOMIZATION.md file in the repo walks through every customization option. The short version:

Must edit: Your personal context file (run the identity scan or fill in .personal/CONTEXT.md manually).

Should edit: The team-ops directory with your team’s real data. The offer ladder template with your pricing structure.

Optional: Add your own skills for processes specific to your business. Customize the confidentiality topic guards in the prompt-submit hook. Adjust the overnight automation prompts for your priorities.

The canon frameworks, quality standards, and core hooks work out of the box. You don’t need to touch them unless you’re adapting the underlying methodology.

Understand Why It Compounds

Most AI tools are session-based. You open a conversation, do some work, close it, and start over. The work product survives, but the process knowledge doesn’t.

The PRISM breaks that pattern in three specific ways.

First, session logs capture what happened, what worked, what didn’t, and what was learned. Those logs are structured data, not just chat history. Scripts can analyze them for patterns, cost tracking, and optimization opportunities.

Second, SOPs update themselves. When Claude discovers a better approach during a session, the SOP gets a version bump, a new learning entry, and an updated date. The next time anyone on your team encounters that scenario, the improved process is already in place.

Third, the identity layer evolves. Refresh the memory bank quarterly, and the system absorbs changes in your business: new clients, new team members, new priorities, shifted strategy. Claude doesn’t drift out of date because the refresh process is built into the system.

This is the difference between using an AI tool and operating an AI system. Tools do what you tell them. Systems learn what you need.

Get Started Today

PRISM is open source on GitHub. Clone it, run setup, run the identity scan, and start working. The entire installation takes less than five minutes. The identity scan takes another 10 to 20 minutes, depending on how many sources you connect.

If you run a content agency, a consulting practice, a coaching business, or any founder-led company where you’re the bottleneck, the PRISM is built for you. It turns Claude Code from an assistant who forgets everything into an operator who remembers everything and gets better at your work every single day.

• Category: Uncategorized