Six boring rules that make AI coding actually work

Executive overview

Most non-technical builders waste hours with AI coding tools because they skip the basics. Six simple rules — applied consistently — deliver 90% of the results without the complexity of sub-agents, MCP servers, or custom tooling.

The rules work together: plan first, manage context deliberately, accumulate lessons, specify the what not the how, and let tests close the loop.

The core insight: AI intelligence degrades as context fills — structure your workflow around that constraint.

Plan before you build

• Use AI to interview you, one question at a time, before writing a single line of code
• At the end of the interview, ask the AI to save the plan as a markdown file
• As work progresses, keep the plan updated — mark features done, note what's next
• Starting each new conversation, point the AI to the plan so it knows where it stands
• Generic built-in plans are not enough; the interview produces a plan specific to what you want

Restarting when a feature fails

• If the AI fails on a feature two or three times in a row, stop trying in that thread
• Before starting fresh: ask the AI to summarise all errors and failed attempts in a dense message
• Undo the AI's changes back to before it started building the broken feature
• Open a fresh conversation, paste the error summary, and ask the AI to update the plan with those lessons
• The updated plan tells the next conversation what not to do when rebuilding that feature

Start fresh conversations constantly

• Every finished feature or fix is a trigger to start a new conversation
• Carrying unrelated context from a previous feature pollutes the next one
• Two triggers: AI is making repeated mistakes, or you are starting something new
• After completing a piece of work, ask the AI to mark it done in the plan, then open a new thread

Save lessons to your instructions file

• Every tool has a system instructions file: CLAUDE.md, agents.md, cursor rules, etc.
• Two failure modes: ignoring the file entirely, or filling it with irrelevant noise
• Add only lessons the AI needs to avoid repeating past mistakes in this project
• Key cases: new APIs or models released after the AI's training cutoff, recurring bugs that took effort to fix
• Over time this creates compounding engineering — the scaffolding grows and the AI gets more reliable, not less

Be specific about what, not how

• Vague prompts produce vague results — the AI guessing wrong is your fault, not the AI's
• Specify: location, behaviour on interaction, edge cases, confirmation flows — not implementation
• Dictation naturally produces more detail than typing; use a dictation tool to write prompts

Give the AI a way to check its own work

• Instruct the AI in the plan to write tests before building each feature
• The test defines the pass condition; the AI then writes the feature and runs the test
• On failure, the AI reads its own errors and iterates until the test passes
• This self-healing loop lets the AI run autonomously without needing constant human correction
• Once a feature passes its test, hand it back to yourself for final validation