Code 231, Kenitra (2026)

01/09/2026

Basic prompting is not the same as using ChatGPT well.

These 11 moves show the difference.

Prompting is only one part of the system.

The quality of the result also depends on:

→ Which mode you choose
→ Which model handles the tas
→ Where the context lives
→ What source material you provide
→ Whether the information is current
→ Which tools AI can access
→ How clearly success is defined

The biggest mistake is expecting original work from generic inputs.

When building a brand, give AI your real assets, screenshots, notes, and examples.

Those materials shape the result more than another paragraph of prompt instructions.

Then save successful repeated work as a Skill.

Codex can use that Skill to complete tasks across your computer, browser, and other tools.

The workflow becomes more capable each time you turn a proven process into something reusable.

ChatGPT is only a search box if that is how you choose to use it.

-----

➕ Follow us for practical ways to improve AI output.

01/09/2026

This is what the enterprise AI stack looks like in 2026.

A few years ago, the conversation was mostly about which model to use.

Now, that feels like only a small part of the picture.

Once you start thinking about production AI, you quickly run into everything around the model: agents, workflows, search, RAG, observability, security, governance, and the applications people actually use every day.

That is why I like looking at the ecosystem in layers:

→ AI Platforms - Microsoft Foundry, Vertex AI, Bedrock, watsonx, Databricks
→ Enterprise Copilots - Microsoft 365 Copilot, ChatGPT Enterprise, Claude Enterprise, Gemini, Glean
→ Agent Platforms - Copilot Studio, Agentforce, ServiceNow AI Agents, AgentCore
→ Workflow Automation - Power Automate, UiPath, Workato, n8n, Zapier
→ Search & Knowledge - Glean, Elastic, Coveo, Sinequa
→ RAG & Vector Infrastructure - Pinecone, Weaviate, MongoDB, Neo4j, Qdrant
→ AI Coding - GitHub Copilot, Codex, Claude Code, Cursor, Amazon Q
→ Observability - LangSmith, Arize AI, Langfuse, Fiddler, Galileo
→ Security & Governance - Purview, Cisco AI Defense, Lakera, Protect AI, Credo AI
→ Enterprise AI Apps - Writer, Moveworks, Harvey, Sierra, Aisera

The part that stands out to me is this:

The model is no longer the architecture.

The architecture is everything you build around it to make AI useful, secure, observable, and reliable in the real world.

That is where enterprise AI gets interesting.

Which layer of this ecosystem are you spending the most time on right now?

01/09/2026

Claude improves when you stop treating every request like a one-shot chat.

__________

Claude Hacks Nobody Told Beginners starts with context, not clever wording.
Do not copy one-shot ChatGPT habits into a context-rich Claude workflow.
Build a working relationship around repeatable standards.

→ 1. Build a conversation
Do not restart the reasoning with every message.
Carry decisions, corrections, and definitions forward intentionally.

→ 2. Give Claude a role and goal
Use “Act As” to define responsibility, audience, and perspective.
Add the outcome and decision the work must support.

→ 3. Be specific
State inputs, constraints, evidence, format, length, and deadline.
Tell Claude what to avoid and how success will be checked.

→ 4. Use Projects for recurring work
Create separate Projects for content, sales, research, or client delivery.
Keep approved instructions, examples, and reference files together.

→ 5. Upload the actual files
Give Claude the source document, spreadsheet, brief, or transcript.
Use Memory for durable preferences, not temporary project noise.

→ 6. Provide 2–3 strong examples
Show the voice, structure, and quality level you want repeated.
Explain why each example works instead of expecting imitation by guesswork.

→ 7. Ask for critique before creation
Have Claude identify gaps, contradictions, weak evidence, and missing context.
Fix the brief before generating a polished answer to the wrong problem.

→ 8. Let Claude interview you
Ask it to question you until it can build a complete prompt.
Answer with facts, trade-offs, constraints, and unacceptable outcomes.

→ 9. Stack prompts
Separate context, analysis, draft, critique, revision, and final formatting.
Add a checkpoint between stages where human judgment matters.

→ 10. Use Claude as a teacher
Ask “Explain Like I’m 12” when foundations are unclear.
Then request examples, questions, and a short test of understanding.

→ 11. Use Claude to think
Ask for alternatives, assumptions, risks, and the strongest objection.
Use it to improve decisions, not merely produce more words.

→ 12. Transform one output into 10 assets
Turn an approved source into emails, posts, FAQs, scripts, and sales material.
Preserve the original claims and adapt the format for each audience.

For founders, save time without outsourcing creativity or accountability.
Claude can accelerate the method.
You still choose the insight, standard, evidence, and final decision.

Use one repeatable loop:
Context → output → critique → improve → transform.
Save the winning sequence as a template inside Projects.

Test it on one real workflow this week.
Measure editing time, factual corrections, consistency, and usefulness.

Which Claude hack would remove the most rework from your process first?

31/08/2026

50 million lines of code. 1 day. Zero human engineers.

Original Post Below

⬇️ ⬇️ ⬇️

50 million lines of code. 1 day. Zero human engineers.

Claude Fable 5 just redefined what AI agents can do at scale….

Most people don't understand what this model actually does. Let me break it down.

Language models are the engine powering next-generation AI agents and Fable 5 just changed what that engine is capable of.

📌 First, what is Claude Fable 5?

Think of Anthropic's models like floors in a building.

Haiku → Ground floor. Fast, lightweight, cheap.

Sonnet → Middle floor. Balanced, reliable.

Opus → Top floor. Powerful, complex reasoning.

Fable 5 → The penthouse. A brand new floor above Opus that didn't exist before. Anthropic calls it the Mythos-class tier.

📌 What does it actually do differently for AI agents?

1. It works longer without losing focus

- Previous models drift or make errors on very long tasks
- Fable 5 stays sharp across millions of tokens - think acting on an entire company's worth of documents without forgetting the beginning

2. It checks its own work

- Most AI models generate an answer and stop
- Fable 5 reviews what it produced, spots its own mistakes, and fixes them before you even see the output

3. It handles real engineering work at scale

- Stripe had 50 million lines of code that needed updating across their entire system
- A full engineering team: 2+ months
- Fable 5: done autonomously in a single day
- That's not a writing assistant. That's an AI agent acting like a full engineering team.

4. It sees and understands visuals

- Give it a screenshot of a website and it rebuilds the source code from scratch
- No instructions needed — just the image

5. It remembers and improves mid-task

- Fable 5 takes its own notes during long tasks and uses them to get better as it works
- Like a human who writes things down ,except it gets smarter with every note

📌 What stops it from being misused?

Fable 5 has a built-in safety layer called a classifier. Think of it as a security guard running alongside the model.

If someone asks something dangerous , the classifier intercepts it and routes the response to Opus 4.8 instead. The user still gets a helpful answer. The dangerous capability never activates.

Less than 5% of sessions ever trigger this. For everyone else, Fable 5 runs at full power.

1,000+ hours of external red-teaming tried to break through. Nobody found a universal bypass.

The bottom line: Fable 5 is the first AI agent model trusted to take on work that previously required a full human team and finish it faster than anyone expected.

50 million lines of code. One AI agent. One day.

Opus or Fable 5. Which would you trust for enterprise AI agents today? 👇

31/08/2026

Claude gets far more useful when you build the right environment around it.

I would not think of this as “24 things to install.”

I’d think of it as 3 layers that make Claude better at real work.

First: Plugins

These can help with debugging, planning, code review, refactoring, security checks, commit workflows, and development setup.

Then: Skills

This is where Claude becomes more capable across repeatable tasks.

→ Frontend design
→ Skill creation
→ MCP server building
→ Web app testing
→ Interactive web artifacts
→ Spreadsheet work
→ Presentation creation
→ PDF processing

And then comes the part I find most useful: MCP Servers.

This is what connects Claude to the tools and context your workflow already depends on.

GitHub MCP can bring in repositories, issues, and pull requests.

Context7 can provide current documentation.

Playwright MCP can support browser testing.

Supabase MCP can connect database and backend operations.

Linear MCP can bring in project and issue context.

Notion MCP can connect knowledge.

Sentry MCP can help investigate production failures.

Figma MCP can bring design context directly into the workflow.

That changes how you use Claude.

Instead of constantly copying context between tools, Claude can work closer to the systems where the actual work lives.

The goal is not to install all 24.

It is to build the combination that removes the most friction from your workflow.

Which would you add first: Plugins, Skills, or MCP Servers?

31/08/2026

Most AI strategy fails because teams jump to tools before choosing the job.

__________

The useful question is not which model is popular this week.
It is what kind of work you are asking AI to do.

The source framework is simple:
Machine Learning.
Deep Learning.
Generative AI.
AI Agents.
Agentic AI.

Each layer solves a different business problem.
Treating them as one vague AI bucket creates bad projects,
bad budgets,
and bad expectations.

Machine Learning is the analyst layer.
1. Revenue and demand forecasting turns past data into planning signals.
2. Risk and fraud detection looks for unusual patterns before losses grow.
3. Predictive maintenance spots failure risk before operations stall.

Deep Learning is the perception layer.
4. Visual inspection and analysis helps teams review images and video.
5. Speech and voice processing converts spoken work into usable workflows.
6. Document understanding extracts meaning from invoices, forms, and claims.

Generative AI is the creator layer.
7. Knowledge assistants and summarization make policies and meetings easier to use.
8. Content, report, and code drafting speeds up first drafts and internal output.
9. RAG and business knowledge answers connect generated responses to company data.

AI Agents are the manager layer.
10. IT service desk agents triage tickets, diagnostics, resets, and support work.
11. Finance operations agents validate invoices, match records, and trigger approvals.
12. Sales operations agents research accounts, update CRM data, and coordinate next steps.

Agentic AI is the orchestration layer.
13. Procurement orchestration coordinates sourcing, vendors, approvals, and orders.
14. Claims or loan processing connects intake, validation, decisions, and escalation.
15. Incident response orchestration coordinates detection, diagnosis, remediation, communication, and reporting.

This is the practical takeaway:
start with the job,
then choose the AI layer.

A forecasting problem does not need the same architecture as a support agent.
A document extraction workflow is not the same as an end-to-end procurement system.
A RAG assistant is useful,
but it is not automatically an autonomous operation.

The winning teams will not ask,
“Where can we use AI?”
They will ask,
“Which work layer are we trying to improve?”

That one question removes noise.
It also makes the next AI investment much easier to defend.

31/08/2026

→ 𝗧𝗵𝗲 𝗛𝗶𝗱𝗱𝗲𝗻 𝗪𝗼𝗿𝗹𝗱 𝗼𝗳 𝗔𝗣𝗜 𝗧𝗲𝘀𝘁𝗶𝗻𝗴: 𝗔𝗿𝗲 𝗬𝗼𝘂 𝗖𝗼𝘃𝗲𝗿𝗶𝗻𝗴 𝗔𝗹𝗹 𝗕𝗮𝘀𝗲𝘀?

APIs are the silent engines powering modern applications. But how well do we really test them?

• Load Testing: Simulate real-world traffic to see if your API holds steady under pressure. Will it deliver consistent performance when millions rely on it?

• Smoke Testing: The quick health check. Does your API start and respond without critical errors? It’s the first line of defense.

• Functional Testing: Verify each endpoint behaves exactly as expected. Are all features working seamlessly, or are hidden bugs lurking?

• Integration Testing: Your API rarely works alone. Does it play well with databases, third-party services, and other components? Integration is where many projects fail.

• Stress Testing: Push your API beyond normal limits. When chaos strikes, will your system gracefully degrade or crash?

• Regression Testing: After every fix and update, does your API retain its reliability, or have you introduced new faults?

API testing isn’t just a checkbox - it’s the foundation of trustworthy digital experiences.

--------------

If AI still feels confusing
it is not your fault.
The problem is scattered tools and no clear roadmap.
If you want a structured path from basics to production AI
👉 Comment/DM 𝗔𝗜

Follow us for more insights

31/08/2026

An AI policy tells people what is allowed. AI governance proves who is accountable.

__________

Most leadership teams now have an AI policy.
That is useful.
It is also not the same as AI governance.

A policy usually answers usage questions.
What tools can employees use?
What data is off limits?
Who needs approval?
Where are the boundaries?

Those rules matter because they reduce confusion.
They help employees understand acceptable use.
They make training and compliance easier to document.
They create a baseline for responsible behavior.

But a policy is still mostly a document.
Governance is the operating system around that document.

Governance asks harder board-level questions.
Who approved this AI system?
Who owns the risk?
Who monitors performance?
Who checks for harm, bias, or failure?
Who reports back when something changes?

That difference becomes critical once AI moves from experiments into decisions.
A chatbot policy may help staff avoid unsafe prompts.
It will not, by itself, decide who owns a model that affects customers.
It will not prove that high-risk systems are monitored.
It will not give the board a clear line of sight.

AI policy usually lives in handbooks, intranets, and compliance packs.
AI governance lives in board agendas, risk registers, governance forums, and operating models.
It also needs regular review.

Policy shows up when people know the rules.
Governance shows up when decisions are traceable.
Policy is measured by approval, training, and acknowledgement.
Governance is measured by active oversight, named ownership, and assurance.

The leadership mistake is treating the first as a substitute for the second.
That creates a false sense of readiness.
It looks organized on paper.
But it can still leave no one clearly accountable when an AI system causes damage.

The stronger question is not:
Do we have an AI policy?

The stronger question is:
Can we prove who is accountable for each meaningful AI decision?

That is where AI governance starts.
And it is the part boards will be expected to understand.

31/08/2026

GraphRAG is what happens when RAG stops treating documents as disconnected pieces.

The classic RAG process usually works like this:

→ User asks a question
→ The question is converted into an embedding
→ The system retrieves matching chunks
→ The retrieved context is sent to a Large Language Model
→ The model generates an answer

This works well for direct questions.

But enterprise knowledge is rarely that simple.

Important answers are often spread across multiple documents and connected through people, entities, claims, teams, projects, customers, risks, and timelines.

That is where GraphRAG comes in.

It adds a knowledge graph layer before the question is even asked.

The first stage happens offline.

Documents are broken into smaller text units.

The system extracts entities, relationships, and claims.

It builds a map of how everything is connected.

It identifies communities of related information and summarizes them.

It generates embeddings for the text units, graph descriptions, and summaries.

This creates multiple knowledge stores:

→ Original documents
→ Smaller text units
→ Graph tables
→ Community summaries
→ Vector index

The second stage happens at query time.

The system can use different search paths to find the answer:

→ Local search for entity-specific questions
→ Global search for broad dataset-level questions
→ DRIFT search for exploratory investigation
→ Basic vector search for similarity-based retrieval

Instead of only asking:

“Which chunks are closest to this question?”

GraphRAG can ask:

→ What entities are involved?
→ How are they connected?
→ Which communities summarize this topic?
→ Which source passages support the answer?
→ What relationships explain the pattern?

This is why GraphRAG matters for enterprise intelligence.

It helps AI move from text retrieval to knowledge understanding.

Better answers do not come only from larger context windows.

They come from better structure, better retrieval, clearer provenance, and a deeper understanding of how information connects.

GraphRAG is not meant to replace RAG.

It is an upgrade for questions that require relationships, synthesis, and reasoning across connected knowledge.

Which use case do you think needs GraphRAG the most: compliance, research, customer intelligence, risk analysis, or enterprise search?

Follow us for more such insights!!

30/08/2026

AI PM Cheat Sheet

Original post:
__________

Most people learning AI Product Management are learning it backwards.

They start with models, prompts, RAG, agents, and whatever AI framework is trending that week.

But that's not where I would start.

If I were evaluating an AI PM today, I'd want to know how they think when the technology gets messy.

Can you identify a problem that actually needs AI?
Can you define what a "good" output looks like?
Can you decide when the model should act and when a human should take over?
Can you make the tradeoff between accuracy, latency, cost, and user experience?

And most importantly: can you tell whether the product is actually working?

That's the part of AI Product Management that gets overlooked.
A good AI PM needs to think across the entire system:
Problem discovery before model selection.
Data strategy before assuming the model will figure it out.

Evaluation before launch.
Human-in-the-loop design before something goes wrong.
Metrics beyond clicks and usage.

Safety and trust as product decisions, not compliance tasks at the end.
And continuous iteration because AI products don't behave like traditional software.

The model is only one part of the product.
The experience around it is what users actually judge.

I've put together this AI PM Cheat Sheet covering the areas I think matter most: discovery, solution design, data, models, experience, evaluation, GTM, metrics, safety, and iteration.

If you're trying to move into AI Product Management, don't try to learn every AI tool available.

Learn how to make good product decisions when AI is part of the system.
Tools will change.
That skill won't.

Code 231

01/09/2026

01/09/2026

01/09/2026

31/08/2026

31/08/2026

31/08/2026

31/08/2026

31/08/2026

31/08/2026

30/08/2026

Address

Website

Alerts

Shortcuts

Share

Category

Listing takedown request