Agentic Engineering Tips and Tricks: How to use AI to Build Faster
Building solo with limited time, I've found a few AI workflows that genuinely move the needle:
- It's cheaper to refine a plan than fix a broken implementation.
- Use the best models because time is more valuable than tokens.
- Write cursor rules so the AI follows your conventions.
- Use Wispr Flow to transcribe your speech rather than typing.
- Keep your architecture clean from the start, because the AI replicates what it sees in your code.
- Use ChatGPT and Claude free plans for small, isolated tasks to keep costs down.
Building things solo, bootstrapped, and next to a job means that my time is always scarce. So I've relied heavily on AI as a sparring partner and coding assistant while building postfern.
Over the past weeks, I've tried a variety of AI tools and workflows (including Cursor, Claude, Wispr Flow, and GitHub Copilot) and have found some things that boost my productivity.
Below is a list of tips that help me get more done in less time in the cheapest way possible.
Smart Editor
The landscape for coding agents is evolving so quickly that it's hard to keep up. It seems like every week the coding community crowns a new best-in-class tool. At the moment, the three most popular tools are: Cursor, Claude, and Codex.
I prefer Cursor over the other two because:
- Cursor is a fork of VS Code, so it feels familiar.
- Unlike Claude, Cursor lets you choose from a range of model providers.
- I prefer working with a sidebar interface, rather than from the command-line like Claude.
- Codex emphasizes multi agent coding, which I don't do because I like to run and check in small steps.
Sidenote: I've curated a list of really helpful tools for building in my other post. If you're looking for helpful tools to ship faster, that list is a good place to start.
Plan Obsessively
It's far cheaper (in time and tokens) to optimize a plan than to repair a lousy implementation built on a scrappy plan. So before letting the AI agent write code, I plan obsessively. For small and medium sized features, I usually start by giving the AI a verbose brain dump of what I want to build and tell it to write a plan in Cursor's plan mode. Then I refine the plan by challenging it with different models.
For larger features and refactors, I start my plan by creating an empty .md
file, and prompt the agent to write a structured implementation plan based on my
input. To the prompt, I like to add that the plan should be concise, written at
a senior engineer level, broken into small incremental steps, and include the
relevant files. To implement the plan, I prompt the agent to proceed in steps so
I can sanity check each step before moving to the next.
Use Top Models
When I build new features or make large refactors, I always use the best models available. Downgrading to a cheaper model leads to worse output and the tokens saved upfront are just spent later on corrections. At the time of writing, the best coding models are: Claude Sonnet 4.5, Claude Opus 4.5, and GPT 5.2. But that information will be outdated quickly.
Write Rules
Without guidance, the AI steers towards spaghetti code that's hard to understand
and requires a lot of refactoring. Cursor lets you create a .cursor/rules
directory in your project root where you can place Markdown files with rules
that the AI should follow. In my rules directory, for example, I define things
like code conventions for React, project file structure, naming conventions, and
available UI components.
Use Dictation
My most recent discovery has been Wispr Flow, which lets me ramble out loud my ideas and have them transcribed. The great thing about Wispr is that it turns your speech into text and types it anywhere you'd normally use a keyboard. So you can use it in pretty much any application. Wispr even cleans up filler words, punctuation, and other noise. I often use it to ramble my prompts out loud because it's much faster than typing.
Clean Architecture
Enforce clean architecture in the project early on, because the AI picks up on the patterns it sees in your code and uses them in its output. So if you begin your project with a well organized monorepo, clear separation of concerns, and well-defined file structure, the AI is more likely to follow those patterns in its output.
Free Plans for Small, Isolated Tasks
Above I mention using the best models generously, and that's especially true for bigger coding tasks. But the cost does add up quickly, so for small, isolated tasks I use the free plans of Claude and ChatGPT. I usually keep two browser windows open, one with the free version of Claude and one with the free version of ChatGPT — and use them for sparring, small clean up tasks, or sanity checking.
Final Thoughts
- Use a smart editor, preferably Cursor.
- Time spent planning beats time spent debugging.
- Use top models generously for big tasks.
- Keep a directory of rules in
.cursor/rules/to guide the AI. - Use Wispr Flow to transcribe your ideas out loud.
- If your architecture is clean, the AI is more likely to follow clean code principles.
- Use free plans for small, isolated tasks to save on costs.
