Using LLMs to Plan, Prompt, and Execute Development Work

This guide shows how AI agents and LLMs can dramatically improve developer productivity through structured workflows. The key insight: LLMs excel at structured thinking, repetition, and documentation — but only when you give them the right scaffolding.

---

Why Most LLM-Assisted Development Fails

Traditional approaches fail because developers:

• Jump straight into "make this code" prompts without planning
• Lose context across sessions and days
• Repeat the same planning work for similar features
• Don't preserve decisions and assumptions anywhere
• Let the LLM make choices without tracking them
• Use the wrong model for the task

The solution isn't better prompts — it's a better workflow and choosing the right tool.

---

Choosing the Right Model

Not all models are equal. Using the wrong model has real consequences:

Wrong Choice	Consequence
Cheap model for complex architecture	Subtle bugs, tech debt, hours of debugging
Expensive model for boilerplate	Wasted budget, no quality improvement
Any model without references	Generic output that doesn't match your patterns

Rule of thumb:

• Flagship models (Claude Opus / Sonnet 4.5, GPT-4o, Gemini 2.0 Pro) → Planning, architecture, debugging, novel problems
• Mid-tier models (Claude Sonnet 3.5, GPT-4o-mini, Gemini 2.0 Flash) → Implementation with clear patterns and references
• Fast models (Claude Haiku, GPT-4.1-mini/nano, Gemini Flash Lite) → Boilerplate, syntax questions, simple tasks

See the Choosing the Right Model page for detailed guidance, cost analysis, and a selection checklist.

---

The Core Workflow

Idea / Requirement
       ↓
LLM creates plan/        ← Use flagship model
       ↓
LLM creates prompts/     ← Use flagship model  
       ↓
LLM scaffolds work-notes/
       ↓
LLM executes tasks       ← Mid-tier usually sufficient
       ↓
Commit after each completed task

Each artifact reduces cognitive load and prevents context loss. The LLM becomes your planning partner, prompt engineer, implementer, and technical scribe.

---

Quick Start Guide

For small tasks (hours to 1-2 days):

1. Create a single PLAN.md
2. Execute with session notes
3. Commit after each logical unit

For medium tasks (1-2 weeks):

1. Create plan/ directory with phase files
2. Generate matching prompts/ files
3. Maintain work-notes/ per phase
4. Commit per phase or sub-task

For large tasks (multi-week / system-level):

1. Break into sub-phases (1a, 1b, 2a, 2b)
2. Keep each file under ~200 lines
3. Consider parallelizing across agents
4. Commit frequently with clear messages

---

The Golden Rules

1. Plan before you prompt — Spend 10 minutes planning to save hours of rework
2. Match model to task — Don't use a cheap model for complex work; the "savings" will cost you
3. Keep session notes — The LLM can't remember what you don't write down
4. Commit after each task — Small commits are reviewable commits
5. Use reference implementations — Show the LLM what "good" looks like in your codebase
6. Phase large work — Break it down until each piece fits in one session

Treat your LLM like a senior engineer who never forgets — but only if you give it structure and pick the right one for the job.

---

Start Here

Quick Reference Card

Security & Sensitive Data

Choosing the Right Model

When LLMs Aren't the Answer

The Workflow

Step 1: Planning

Step 2: Prompt Documents

Step 3: Session Management

Step 4: Execution & Commits

Step 5: Scaling the Workflow

Team & Operations

Team Collaboration Patterns

Common Pitfalls & Troubleshooting

Measuring Success

Worked Examples

Example: Helm Chart

Example: Ansible to Temporal

Resources

Templates & Checklists

Building a Prompt Library