Framework

IEPEV

Intention → Exploration → Planning → Execution → Validation

A structured methodology for human-AI collaboration.
AI accelerates each phase — the human keeps the steering wheel.

Core Principle
AI accelerates everything,
except the intention.
The intention stays 100% human.
01
Human sets the intent

AI handles micro-tasks, but framing the problem, the vision, and the constraints stays with you. Think "big picture".

02
AI drafts, human decides

AI produces first versions and components. The human reviews, validates, and decides. AI never replaces judgment.

03
Always validate the output

AI output is presumed incomplete. Validation is mandatory: logic, integration, edge cases, ethics.

I
Phase 1 of 5
Intention
Define the problem, the vision, the constraints, and provide context

What you do

  • Frame the problem you're solving
  • Set the constraints and boundaries
  • Define what success looks like
  • Think "big picture" — AI fills technical gaps later

Why it matters

This is the only phase that is entirely human. A well-defined intention prevents the AI from drifting into irrelevant solutions. The clearer the intent, the better every subsequent phase.

Responsibility
🧠
100%
Human
🤖
0%
AI
E
Phase 2 of 5
Exploration
Analyze the context, identify resources and constraints

What happens

  • AI scans the codebase, docs, market data
  • Identifies patterns, dependencies, and risks
  • Surfaces relevant prior art and alternatives
  • Human guides direction — AI does the heavy lifting

RAG through project structure

Your folder structure is your RAG. Organized files, documentation (CLAUDE.md, README), and structured knowledge give the AI rich context without a vector database.

Responsibility
🧠
Guide
Human
🤖
Analyze
AI
P
Phase 3 of 5
Planning
Structure the approach, define the steps, identify the components

What happens

  • AI proposes architecture, steps, and components
  • Human validates the approach and priorities
  • Break the project into autonomous blocks
  • Each block is designed, validated, then built separately

Decomposition principle

Never ask for a complete deliverable at once. A prototype becomes: UI component + data flow + API call + tests. Validate each block before moving to the next.

Responsibility
🧠
Validate
Human
🤖
Propose
AI
E
Phase 4 of 5
Execution
Generate the deliverables — code, content, architecture, analysis

What happens

  • AI produces actual deliverables — code, docs, prototypes
  • Human supervises and course-corrects in real-time
  • Iterate block by block, not all at once
  • Validate each block before advancing

The AI-native product flow

Idea → Justification → Stabilization → Product

The product and process co-evolve. AI helps discover the right solution through rapid iteration — a form of AI-assisted systems engineering.

Responsibility
🧠
Supervise
Human
🤖
Produce
AI
V
Phase 5 of 5
Validation
Test, verify integration, identify edge cases, iterate

What happens

  • Test the integrated blocks together
  • Verify coherence — does the experience flow?
  • Check edge cases, logic, and integration points
  • Human decides — AI assists with testing and analysis

Critical vigilance

In sensitive domains (medical, financial, legal), human verification is not optional — it's an ethical requirement. AI expands capabilities but never replaces deep expertise.

Responsibility
🧠
Decide
Human
🤖
Assist
AI

Your folder structure is your RAG

Structured project layout gives the AI deep context — no vector database required.

CLAUDE.md # AI reads this first — rules & decisions README.md # High-level context (human or AI) Makefile # make pdf, make pitch, make clean │ ├── core/ # Source of truth │ ├── main.tex # Document root │ ├── pitch-deck.tex # Beamer slides │ ├── pitch-deck-speech.md # Speech notes │ ├── LLKBusinessPlan.cls # Custom class │ ├── references.bib # Verified citations │ └── sections/ # 1 file = 1 concern │ ├── 01-resume-executif.tex │ ├── 04-concurrence.tex │ ├── 07-modele-economique.tex │ └── 09-projections.tex │ ├── sources/ # Verified research & data │ ├── africa-telehealth-benchmark.md │ ├── competitors/ # Competitive analysis │ └── funding/ # Investor research │ ├── knowledge/ # Domain expertise, curated │ ├── README.md # Index of knowledge files │ └── sources/ # Raw research materials │ └── exports/ # Generated — never edited ├── Business_Plan.pdf └── Pitch_Deck.pdf
📖

CLAUDE.md = System Prompt

Project structure, build commands, conventions, key decisions. The AI reads this first and stays aligned across sessions.

📁

1 File = 1 Concern

Numbered sections, clear naming. AI can read exactly the file it needs without loading the entire project.

🔍

Sources = Verified Knowledge

Separate research from generated content. AI draws from verified data, not its own hallucinations.

🔄

Exports = Output Only

PDFs are generated, never hand-edited. Single source of truth stays in core/.

IEPEV in practice

01 — Intention

Define the task

"I need a pitch deck with 13 slides covering problem, solution, market, team, and financials"
—
02 — Exploration

Scan context

"Here's the business plan, the competitor analysis, and the funding sources"
Reads CLAUDE.md, sections/, sources/, identifies key metrics and structure
03 — Planning

Structure slides

Validates slide order and content allocation per slide
Proposes 13-slide structure with timing, speech notes, and data points
04 — Execution

Generate content

Reviews each slide, adjusts tone and data
Produces LaTeX Beamer code, speech notes, builds PDF
05 — Validation

Test & ship

Rehearses with timer, checks data consistency across all docs
Cross-checks numbers between pitch deck, business plan, and speech notes

At a Glance

PhaseDescriptionHumanAI
Intention Define the problem, vision, constraints, provide context 100% 0%
Exploration Analyze context, identify resources and constraints Guide Analyze
Planning Structure approach, define steps, identify components Validate Propose
Execution Generate deliverables: code, content, architecture Supervise Produce
Validation Test, verify integration, identify edge cases, iterate Decide Assist

Three pitfalls IEPEV prevents

🌀

Context Loss

AI loses the thread over long prompts. Without structure, outputs drift from original intent. You end up re-explaining everything.

→ Each phase reanchors context with clear boundaries
🧩

Inconsistencies

Each part works alone, but they don't fit together. Different prompts produce conflicting outputs, data mismatches, and broken flows.

→ Block-by-block validation catches mismatches early
🙈

Blind Dependence

Accepting AI output without testing it. Shipping a prototype that was never validated against real constraints or edge cases.

→ Mandatory validation before anything ships

AI is a phase accelerator,
not a substitute for human judgment.

IEPEV structures human-AI collaboration so that AI accelerates each phase while keeping strategic control and decision-making where it belongs — with you.

Intention Exploration Planning Execution Validation
← →   or click  •  F for fullscreen