What I Built
Pup is a VS Code extension that treats your prompt like source code — you fill a form, it compiles to whatever format your model prefers.
The insight behind it: LLMs aren't all trained the same way, and the format of your prompt actually matters.
| Model | Preferred Format |
|---|---|
| Claude (Anthropic) | XML tags |
| GPT-4 / GPT-5 (OpenAI) | Markdown or JSON |
| Gemini (Google) | Markdown or structured JSON |
| Cursor / Copilot Chat | Markdown |
Pup compiles the same prompt into all four. One toggle. No rewriting.
How It Works
-
Open the builder —
Cmd/Ctrl+Shift+P→Pup: Open Builder - Pick a preset — Bug Fix, New Feature, Refactor, Code Review, Research, Agentic Task, or Custom
- Fill the sections — Role, Context, Task, Constraints, Output Format
-
Type
@inside any field to fuzzy-search and reference workspace files - Copy or save — export as Markdown, XML, JSON, or plain text
Saved prompts live in .prompts/ as .md files with embedded frontmatter. Right-click → Open Selected .md as Prompt to pick up exactly where you left off.
A Few Details Worth Knowing
Live token counter. A char/word-blended estimator (±10–15%) shows context window usage in real time as you type — no per-model tokenizer required. Switch your target model and the usage bar updates.
Minimize hallucination toggle. Appends an <investigate_before_answering> directive that forces the model to read referenced files before responding. Surprisingly effective.
Caveman mode. Appends a directive that strips filler from the model's reply — roughly 75% fewer output tokens. Great for high-volume iteration.
Zero outbound requests. No API calls inside the extension. No accounts. No telemetry. Your prompts never leave your machine.
Remote-ready. Works in Remote-SSH, Codespaces, dev containers, and virtual workspaces.
Install
# VS Code
code --install-extension alimalik.pup
# Cursor
cursor --install-extension alimalik.pup
Or search Pup in the Extensions panel (Cmd/Ctrl+Shift+X).
What I'm Looking For
This started as a personal workflow fix. I'm shipping it publicly because if I had this problem, others probably do too.
Feedback I'm specifically curious about:
- Which output formats do you actually reach for?
- What role presets are missing from your workflow?
- Edge cases in the token estimator you've hit?
Source is on GitHub. PRs are welcome — new presets especially.
If it saves you time, a ⭐ on the repo helps others find it.
✨ Why Pup?
LLMs respond better to structured prompts — role, context, task, constraints, output format. Most developers freestyle into a chat window, get mediocre output, then iterate forever.
Pup is a form-based prompt compiler. Fill out sections → Pup compiles them into Markdown / JSON / XML / plain text. Save once, reuse forever. No accounts. No telemetry. No AI calls inside the extension.
Important
Zero outbound requests. Zero API keys. Zero telemetry. Your prompts never leave your machine. Pup is a pure compiler — your prompt, your model, your editor.
🎯 Why Format Matters
Different model vendors are trained against different prompt structures. Matching the format their training data used measurably improves output quality…