The Best Way to Prompt AI: Building PromptSmith

TL;DR: I built a lightweight, local tool that turns brief ideas into structured, model-ready prompts—often in XML—showing the best way to prompt AI for clearer, more consistent outputs.

Context: Why This Specific Sub-Project (for the best way to prompt AI)

Audience: People dabbling in AI who want better outputs without getting overly technical.
Pain point: Quick notes don’t translate into well-structured prompts, so models misinterpret or under-deliver.
Success definition: A repeatable way to expand short ideas into high-quality prompts, plus mini-tools for images, programming prompts, and thumbnails.

What I Built

Goal: Turn my scrappy, brief prompts into clear, structured instructions that models follow reliably.

Scope (sub-project only): Prompt expansion and optimisation; XML formatting where helpful; mini-tools for image prompt templates, a program-builder that considers current libraries, and a thumbnail helper.

Stack & Tools

GPT-5 API for prompt expansion and refinement.
XML formatting to add structure when models benefit from schema-like inputs.
Gemini for multimodal bits, e.g., analysing references for thumbnails.
Local app in Microsoft Edge (runs as an app-like tab) for speed, privacy, and focus.

Timeline

<br>

Date	Milestone	Owner	Notes
2025-08-02	Kicked off initial build	Jacques	Set up base prompt improver
Ongoing	Iteration & modular add-ons	Jacques	Program builder & thumbnails

<br>

Process (Step-by-Step)

Map the pain. I listed my common “too-brief” prompts and defined a target structure (sections, parameters, constraints). Trade-off: keep it simple vs. add knobs—kept it simple first.
Integrate GPT-5 early. Docs were thin, so I shipped a minimal path that did expansion and basic constraint checks, then layered features.
Tame XML bleed-through. Early runs saw the model echo internal scaffolding. I separated stages and enforced stricter output boundaries to stop self-prompting.
Add multimodal helper. Gemini analyses rough visual refs and proposes cleaner thumbnail directions.
Package locally. Wrapped it as an Edge app-like tab for quick access and privacy.

Pull-quote: “When a thought’s half-baked, PromptSmith plates it up so the model actually gets it.”

Results

Clarity: Brief notes become structured prompts the model can follow.
Consistency: Structured outputs (often XML) reduce odd model behaviour.
Speed: Faster from idea → usable prompt, especially for programming and thumbnail briefs.

Obstacles & How I Solved Them

Issue: Early GPT-5 documentation gaps.Fix: Ship a minimal slice and iterate in short loops.Why it worked: Smaller surface area for errors and faster feedback.
Issue: XML conversion causing self-referential outputs.Fix: Stage separation and stricter output boundaries.Why it worked: Prevented the model from internalising scaffolding.
Issue: Feature creep vs. time.Fix: Modularise—keep the core improver simple; add tools as optional blocks.Why it worked: Maintains momentum without bloat.

Lessons for AI Dabblers

Structure wins. Even light sections and parameters beat a rambling prompt.
Separate concerns. Generate content first; format/convert second to avoid bleed-through.
Keep it close. A quick local wrapper means you actually use the tool.

Why It Mattered to Me

I think fast and type faster.PromptSmith lets me keep that pace while giving models the structure they need.It’s my safety net for “did I specify enough?” across text, code ideas, and visual briefs.

FAQ

Q: Why use XML for prompts?A: It adds explicit structure—sections and attributes reduce ambiguity and guide model attention.

Q: Can multiple models play nicely in one flow?A: Yes. Use each where it shines, and keep clear boundaries between stages.