Introduction: why 95% of Claude AI users underutilize it
There is a growing gap between users who exploit Claude AI as a simple chatbot and those who turn it into a genuine productivity lever. This gap is not closed by learning to write better prompts. It is closed by fundamentally changing your method.
This method comes from Andrej Karpathy — co-founder of OpenAI, former AI director at Tesla, inventor of the vibe coding concept, and now researcher at Anthropic where he directly contributes to Claude's training. In other words, few people in the world are better positioned to explain how to get the most out of this model.
During a recent conference, Karpathy detailed three precise techniques that allow, according to him, to achieve 10 times higher productivity than average. This article breaks them down and translates them into concrete practices applicable immediately.
Important context
Andrej Karpathy joined Anthropic in early 2025. His perspective on using Claude is that of a practitioner who understands the model's internal architecture. His recommendations are not opinions, they are observations derived from the system's design itself.
The real problem: context, not the prompt
Most users think that the quality of results depends on the quality of the prompt. This is a diagnostic error. The prompt represents only a fraction of what determines the relevance of a response.
The real lever is context: who you are, what you're trying to accomplish, your business constraints, your reference examples. Without this context, even the best-written prompt forces the model to guess — and a model that guesses cannot aim accurately consistently.
Karpathy offers a useful analogy: Claude is comparable to an exceptional intern. Encyclopedic knowledge, impressive execution capacity, but limited contextual judgment and real blind spots. It can write a complete marketing strategy in 30 seconds and get an elementary date calculation wrong.
Beware of the prompt engineering reflex
Faced with a disappointing response, the natural reflex is to rephrase the request. In the vast majority of cases, the problem is not the formulation: it's the lack of sufficient context. Investing time in structuring context is systematically more profitable than optimizing prompt writing.
Technique 1: Building Controlled Memory
The problem of default amnesia
By default, Claude starts from scratch with each new conversation. Without intervention, you are doomed to re-explain your identity, your project, your tone, and your rules with each session. This is a considerable waste of time and a source of inconsistency.
Claude now has an automatic memory feature between conversations. However, Karpathy warns against a major pitfall: this memory escapes your control. It is the model that decides what it retains, what it forgets, and how it restores it. Yet context is precisely your number one lever — delegating it to autopilot amounts to abandoning the most impactful tool in your arsenal.
The solution: Claude Projects
Professionals build a memory they control entirely, through the Projects feature on Claude.ai. A project functions like a dedicated workspace, equipped with permanent instructions that you write yourself.
Recommended structure for project instructions:
- Identity: who you are, your role, your sector
- Project objective: what you're trying to accomplish in this space
- Tone and style: register, writing constraints, prohibitions
- Business rules: constraints specific to your activity
- References: documents, examples, guidelines to integrate
Example of instructions for a business project:
1Context: I am a cybersecurity consultant specializing in SMEs.2My target: CIOs and IT managers of organizations with 50 to 500 employees.3Tone: Expert, direct, no unnecessary jargon. Never condescending.4Default format: Bulleted structure, short sections.5Prohibited: Never cite competing solutions without comparative analysis.6Objective of this project: Produce client deliverables (reports, business proposals, technical summaries).Advanced tip
Feed your projects with reference documents: your best past deliverables, your editorial guidelines, annotated examples. The richer and more precise the memory, the more the model becomes performant on your specific use cases — and the more difficult this advantage becomes to replicate by your competitors.
Technique 2: Replace Vague Requests with Structured Briefs
From prompt to brief: a paradigm shift
A vague request like "write me a report on this subject" places Claude in an improvisation position. It must infer your actual objective, implicit constraints, and success criteria — uncertainties that degrade result quality.
Karpathy goes beyond Claude's native plan mode, which he considers too superficial. He recommends working with a detailed specification (spec) — the equivalent of technical specifications — before any execution.
The 4 components of an effective brief
A structured brief must answer these four questions:
- Real objective: what concrete result are you aiming for? (not the task description, but the expected impact)
- Necessary context: what information must Claude have to avoid guessing?
- Constraints: tone, format, length, scope, explicit prohibitions
- Execution mode: direct result or step-by-step validation?
Comparative example:
| Standard approach | Structured brief approach |
|---|---|
| Write an ad for this apartment. | Objective: maximize qualified viewing requests. |
| Context: family T4 apartment, quiet residential neighborhood, mid-high standing. | |
| Constraints: warm and professional tone, 150 words max, no sensational language. | |
| Mode: propose the headline first for validation, then write the complete ad. |
Let Claude build the brief for you
If writing a brief seems constraining, you can delegate this step to the model itself with the following instruction:
1Before you start, ask me the necessary questions to understand my objective.2Then list the important decisions and have me validate them before executing.This inversion is particularly powerful: you're no longer guessing what needs to be written, it's Claude coming to extract from you the information he needs to achieve the objective.
Tip for IT teams
In a Microsoft 365 context, this structured brief logic applies directly to Microsoft 365 Copilot. System instructions for Copilot Studio agents play exactly the same role as project instructions in Claude: they define the permanent context in which the model operates.
Technique 3: Make Results Verifiable
The founding principle: learning through reinforcement
Karpathy recalls a central principle in how LLMs are trained: models progress when they have a clear verification signal. The more you give Claude a way to validate his own work, the more results improve.
This logic, which operates at the training level, can be reproduced in your daily interactions.
Level 1: Explicit validation criteria
After submitting your brief, systematically add a verification instruction:
1Before submitting the final result, verify point by point2that the following criteria are met:3- [Criterion 1]4- [Criterion 2]5- [Criterion 3]6Indicate for each criterion whether it is validated or not.This approach forces the model to structured review rather than implicit validation.
Level 2: Integrate real data
The next level is to introduce real performance data into the loop. Claude is no longer guessing what works — it relies on your concrete results to calibrate its future productions.
Example of iterative improvement prompt:
1Here are the results from my last 3 productions:2- Production A: [metrics]3- Production B: [metrics]4- Production C: [metrics]5 6Analyze what performed best and identify the patterns.7Use these insights to produce the next version by explicitly relying on what works.The virtuous loop
These three techniques, combined, create a self-improving system:
- Memory capitalizes on your business context
- Brief ensures alignment with your real objectives
- Verification anchors the model in your measured results
Critical point: stay in control
Karpathy insists on this point: don't disappear from the loop. Total automation without human supervision produces results that gradually drift from the real objective. Your role is that of a pilot, not a passenger. It is you who defines what matters, what is true, and what deserves to be built.
Practical Implementation with PowerShell and the Anthropic API
For IT teams wishing to integrate these principles into automated workflows, here is an example of a PowerShell script that structures a request to the Claude API with the three components (permanent context, structured brief, verification criteria):
1# Prerequisite: Anthropic API key in ANTHROPIC_API_KEY environment variable2# API Documentation: https://docs.anthropic.com/en/api/getting-started3 4$ApiKey = $env:ANTHROPIC_API_KEY5$ApiEndpoint = "https://api.anthropic.com/v1/messages"6 7# Permanent context (equivalent to project instructions)8$SystemPrompt = @"9Context: You assist a Microsoft 365 administrator in writing technical documentation.10Tone: Technical, precise, structured. Short sentences. Bulleted lists preferred.11Format: Sections with H2/H3 headings, code blocks for all commands.12Prohibited: Never assume a default configuration without flagging it explicitly.13"@14 15# Structured brief (objective + context + constraints + execution mode)16$UserBrief = @"17Objective: Write a procedure for deploying password policy via Entra ID.18Context: Hybrid environment Azure AD Connect + Entra ID. 500 users. M365 Business Premium tenant.19Constraints: Include PowerShell commands (Microsoft.Graph module). Max 600 words. Target audience: junior administrators.20Validation criteria:21- All commands are testable and syntactically correct22- Prerequisites are listed at the beginning of the document23- Each step includes a result verification24Mode: Propose the procedure outline first for validation, then write the complete document.25"@26 27$Body = @{28 model = "claude-opus-4-5"29 max_tokens = 204830 system = $SystemPrompt31 messages = @(32 @{33 role = "user"34 content = $UserBrief35 }36 )37} | ConvertTo-Json -Depth 538 39$Headers = @{40 "x-api-key" = $ApiKey41 "anthropic-version" = "2023-06-01"42 "content-type" = "application/json"43}44 45try {46 $Response = Invoke-RestMethod -Uri $ApiEndpoint -Method Post -Headers $Headers -Body $Body47 $Response.content[0].text48} catch {49 Write-Error "Anthropic API error: $($_.Exception.Message)"50}Official references
- Anthropic API Documentation: https://docs.anthropic.com/en/api/getting-started
- Managing Claude.ai Projects: https://support.anthropic.com/en/articles/9517075-what-are-projects
- Microsoft Graph PowerShell Module: https://learn.microsoft.com/en-us/powershell/microsoftgraph/
Summary: the 3 techniques in practice
Build controlled memory
Create a dedicated Project in Claude.ai for each recurring work domain. Write permanent instructions covering your identity, objectives, tone, and business rules. Feed it with your reference documents. Control yourself what the model knows — don't delegate this decision to automatic memory.
Structure briefs with 4 components
Before any significant request, formulate a brief answering these four questions:
- What is my real objective (expected impact, not task description)?
- What context must Claude have to avoid guessing?
- What are my constraints (format, tone, length, prohibitions)?
- What execution mode: direct result or step-by-step validation?
If brief writing is complex, ask Claude to ask you the necessary questions before executing.
Integrate verification criteria
Systematically add a list of explicit validation criteria to your briefs. Ask the model to verify point by point before submitting the result. For recurring use cases, reinject your real performance data to create a continuous improvement loop. Stay involved in each cycle — you are the pilot, not the passenger.
Conclusion: Become a better pilot, not a prompt expert
Karpathy's three techniques have a common denominator: at each step, you decide. You define the context, you frame the objective, you establish success criteria.
Claude can write, analyze, and execute for you. It cannot understand for you what really matters, what is accurate in your specific context, and what deserves to be built.
The goal is not to master prompt engineering. It is to become a more precise and demanding pilot of your AI — whether it's Claude, Microsoft 365 Copilot, or any other language model integrated into your work environment.
These reflexes require initial investment, but they constitute precisely the barrier that separates those for whom AI remains a gadget from those who transform it into a sustainable competitive advantage.
