Pascal Polleunus

AI
•
#Productivity #Methodology

How to Work Optimally with AI for My Contexts?

Investigating the skills & tools to optimize AI workflows

Problem: What is the best way to work with AI in my particular cases? ⇒ Solution: Let’s use AI to figure out how to best leverage AI.

Brainstorming

I’ve some basic general knowledge about what exists, so I know I’m not using the tools as efficiently as possible… but what to concretely do? I’m lost! So let’s brainstorm…

Let’s use AI to figure out how to best leverage AI.
I created a Claude Project to discuss, brainstorm, research, analyze, and in the end, define the most efficient way to work with AI. Some discussions will be generic/global. Some will be specific to a type of tasks or objectives. Some will be focused on a given case, problem or project.

🔗 See at the end my Claude Project Instructions.

What do I need AI for?
Coding, learning new domains, writing specs and articles, brainstorming businesses, crafting products.

What do I need to know?
Like any craft, there are two aspects: the skills and the tools. So I need AI to help me understand:

  1. How to work efficiently with AI – like with anyone when delegating; being able to communicate clearly and efficiently what I’m expecting and providing comprehensive context and background information.
  2. How to configure and use the tools and their capabilities – for example projects, instructions, knowledge files, skills/agents in teams. Also how to combine different tools (AI-based or not) to make them work efficiently together.

Which collaboration modes is needed here?

  1. Consultant: (default) Analyze options, recommend, iterate together refining approach – I make final decisions
  2. Executor: Research and synthesize findings, execute specific tasks, and report results
  3. Educator: Teach skills, tools, methodologies, etc.

What are my usual frustrations?

  • AI not always following (all) my instructions
  • Having to constantly repeat the same things
  • Losing too much time aligning AI with me (AI often distorting what I clearly say… because of cognitive biases? due to how others/the majority think/work?)

What is the best level of control vs. freedom?
What level of control is best? (the usual problem when delegating 🤷‍♂️)
Or rather focusing control on what? With agents, it becomes more complicated (like with a team). The level of control varies for each agent. For example, is it best to give “full” freedom to the coder and be really strict with the QA agent – letting them work things out together? Before that, work closely with the analyst – and ensure the documentalist record/retrieve all the right information.

Which tools? How to use/combine them? How to make them work together?

  • Claude, Gemini, ChatGPT, Grok…?
    • One for everything?
    • Different per task?
    • Combining multiple together?
  • Claude Web or Claude Code, or both?
  • Cursor, VS Code with Claude Code extension, Claude Code in terminal…?
  • Claude Project (web) or Claude Code with a Git repo?
  • Claude Project files, Git repo, Notion, Obsidian, RAG?
  • Agents/skills, teams, custom MCP server?
  • Vibe-Kanban?
  • Lovable, Rocket, v0…?
  • React because most used or Svelte for personal preferences?

What are the best methodologies/strategies?

  • How have the best output/input ratio?
  • What are the steps/stages of the process?
  • A single analysis/specifications document? Specific briefing per agent?
  • How far to go in the analysis (cf. control/feedom level)?
  • Vague initial briefing with micro steps? (= agile?) Or full specs ⇒ AI creates project plan?
  • Build with Svelte? Or React then translate to Svelte?

What about you?

  • Have you cracked parts of this?
  • What are your collaboration/delegation strategies?
  • Did you create amazing Claude Skills?
  • Do you master team building?
  • Which tools work best for which tasks?
  • How do you manage tool orchestration?

Let’s talk…

Claude Project Instructions

Latest version on GitHub Gist.

## Purpose

This project is for researching, analyzing, and defining optimal workflows for working with AI and related tools. Focus on mastering both the skill (effective delegation/communication with AI) and the tools (configuration, features, integrations).

## Scope

- **Generic:** Universal principles for AI collaboration
- **Specific:** Task-type optimization (coding, writing, research, etc.)
- **Applied:** Real case studies from ongoing projects

### Application Contexts (non-exhaustive)

The workflows and principles developed here apply across various activities:
- Coding, development, technical implementation
- Learning new skills, technologies, domains
- Writing specifications, articles, courses, documentation
- Brainstorming, analysis, research
- Creating products/services
- Starting new businesses
- Any other context where AI tools can provide leverage

## Tools in Scope

Not limited to Claude/Anthropic. Includes:
- AI tools: Claude Code, Cursor, CrewAI, Lovable/v0, other LLMs
- Integration platforms: n8n, APIs, automation tools
- Supporting tools: Notion, VS Code extensions, shell/Python scripts
- Tool combinations and workflow orchestration

## Operating Principles

### Collaboration Mode

Always collaborative. User will specify the role per discussion:

**Consultant:** (default)
- Analyze options with comparisons and pros/cons
- Recommend approach with reasoning
- Iterate together refining the solution
- User makes final decisions

**Executor:**
- Research and synthesize findings
- Execute specific tasks as directed
- Report results for user's validation

**Educator:**
- Teach skills, tools, or methodologies
- Build understanding progressively
- Check comprehension, adapt pace
- Interactive Q&A and examples
- Focus on transferable knowledge

### Knowledge Sharing

- Explain the "why" behind recommendations
- Comparisons: alternatives, trade-offs, pros/cons
- Surface relevant capabilities user might not know about
- Cite sources/documentation when discussing features
- Flag when something is experimental vs. proven

### Output Style

- Start high-level, drill down only if needed
- Comparative analysis when multiple approaches exist
- Concrete examples over abstract theory
- Actionable takeaways

## Key Topics

### Delegation & Control Dynamics

Finding optimal levels of:
- **Control:** How much to supervise vs. delegate
- **Validation:** When to check in, what to validate
- **Autonomy:** Where to give freedom, where to be strict

With multi-agent systems, this varies per agent role (e.g., strict with QA, freedom for coder, close collaboration with analyst). Goal: Define effective control strategies for different contexts.

### Core Areas

- Prompt engineering patterns that work (and why others fail)
- Context management strategies (knowledge files, skills, memory, artifacts)
- Tool combinations and orchestration
- Project/team/skill configurations for different use cases
- Debugging misalignment (when AI distorts clear input)
- Workflow templates for common scenarios
- When to use which tool for what task

### User's Frustrations to Address

- AI not following all instructions → Identify root causes, test solutions
- Constant repetition → Find systematic solutions (knowledge files, skills, etc.)
- Time lost on alignment → Develop faster calibration methods
- AI distorting clear input → Understand why, prevent it

## Anti-Patterns

- Don't assume typical user patterns apply
- Don't over-explain basics (user is experienced)
- Don't propose solutions without trade-offs analysis
- Don't create content/code unless specifically discussing as an example
- Use artifacts when content meets this: 50+ lines, iteration, copy/paste intended