A structured constraint framework for Claude Code in research workflows.
Website | Documentation | Vibe Coding Articles
Without structure, an unconstrained LLM produces the mean of its training data. Ask it to "implement RVQ" and you get a plausible-looking average — not the one matching your paper, your architecture, your constraints. The output compiles, but embeds wrong assumptions, silent numerical bugs, and design decisions made without asking.
The fix isn't better prompts — it's structured constraints. Propel enforces human-in-the-loop gates, domain-specific auditors, and investigation-first methodology so the output goes from "plausible average" to precisely what you need.
| Principle | Rule |
|---|---|
| Assistant, not agent | Investigate and present — don't guess and act. Every claim traceable to evidence. |
| Evidence over agreement | Be correct, not agreeable. Steel-man the counterargument before agreeing. |
| Context discipline | Hallucination risk grows with context. Preserve state in living READMEs, clear proactively. |
| Critical self-reflection | Question your own reasoning as hard as the user's. |
| Break logic loops | Name circular reasoning, reframe, or bring new data. 3-strike limit. |
| Mode | Active Gates | When to Use |
|---|---|---|
| Researcher | Gate 0, 1 | Understanding the problem space — papers, code tracing, approaches |
| Engineer | All (0–4) | Full pipeline from investigation through implementation (default) |
| Debugger | Gate 0, 1, 4 | Root-cause analysis — classify bugs vs. design issues with evidence |
| Trainer | Gate 4 (runtime) | Launch training, monitor, fix CUDA/OOM/path errors |
Switch anytime: /switch researcher, /switch engineer, /switch debugger, /switch trainer.
git clone https://github.com/KevinBian107/propel.git
cd propel && pip install -e .
cd /path/to/your/project
propel initThen start Claude and run /intro to select a mode and set up your project. See the Getting Started guide for details.
Full documentation is available on the Propel website:
- Getting Started — Installation and first workflow
- Core Principles — The five non-negotiable principles
- Pipeline — Gates, questioners, and phase transitions
- Modes — Researcher, Engineer, Debugger, Trainer
- Skills — 17 specialized skills by workflow phase
- Agents — 8 domain-specific auditors
- Customization — Project-specific agents, skills, commands
- Common Pitfalls — Known failure modes and anti-patterns
Propel combines ideas from: obra/superpowers, scott-yj-yang/new-prompt, Talmo's sleap-io, Sionic AI, brunoasm's claude skills, Weizhena's Deep-Research, and Context Engineering Template.
External plugins bridged by Propel skills: openai/codex-plugin-cc (via /c-codex), anthropics/claude-code code-review plugin (via /c-review).
Skills vendored directly: anthropics/skills — frontend-design (see source repo for LICENSE).
MIT — Built by Kaiwen Bian and Yuer Tang.