Introduction
Prompt engineering was a hot topic a few years ago, when even small tweaks—like changing the format or the last sentence of a prompt—could produce vastly different results from large language models (LLMs). In the years since, LLMs have become significantly more powerful and robust. They are now far less sensitive to surface-level variations in formatting, wording, or ordering within the prompt.
However, effective prompting remains crucial. The skills required have evolved: today's agents demand a different mindset and a more strategic approach to achieve optimal outcomes.
Looking ahead, it's likely that by 2026 the methods for prompting agents will have shifted again. Staying current with model updates is essential to maintain effectiveness and extract maximum value from these tools.
This article shares my personal workflow for prompting coding agents—specifically Claude Code—to efficiently complete engineering tasks. The techniques discussed are general enough to apply to most coding agents, including Codex.
Why Optimize Your Prompting Approach
You might wonder whether optimizing prompts still matters given how capable LLMs have become. To some extent, the models can understand you regardless. But there are still good and bad prompting practices.
Examples of effective prompts:
- Be explicit about what you want.
- Include instructions on how to test the implementation.
- Clarify any ambiguities upfront.
Examples of poor prompts:
- Be vague or overly general.
- Make assumptions that conflict with the existing codebase.
Moreover, prompting is not a one-off action. It is an iterative process of collaborating with the model to create a plan—for instance, asking the model to surface ambiguities, discuss them with you, or specify which assumptions to adopt.
How I Prompt Claude Code
Below are the key techniques I use to get the most out of Claude Code. While these examples use Claude Code, the principles apply broadly to other coding agents as well.
1. Transcribe, Don't Type
I use a transcription tool to convert my spoken thoughts into text input for the agent. My current choice is FluidVoice, a free, locally-run model that is both fast and accurate for English transcription.
The obvious advantage is speed: speaking is typically two to three times faster than typing, depending on your typing speed. Additionally, it reduces hand strain and frees mental bandwidth for content rather than mechanics.
But there is another benefit: I'm more inclined to include full context when speaking. When typing manually, I sometimes skip details to save effort. Transcription encourages me to provide richer context, leading to better results from the agent. For example, I might describe the overall architecture, edge cases, and testing requirements more thoroughly—something I'd be tempted to abbreviate when typing.