Propel

Overview

Claude Code can be a powerful boost to research workflows — if used correctly and carefully. Without structure, an unconstrained LLM produces the mean of its training data: ask it to "implement RVQ" and you get a plausible-looking average of every RVQ implementation it has seen, not the one that matches your paper, your architecture, your constraints. The output compiles, but it's noisy — wrong assumptions baked in, silent numerical bugs, design decisions made without asking.

The fix isn't better prompts — it's structured constraints. When we constrain an LLM with domain-specific rules, verification gates, and forced checkpoints, the output goes from "plausible average" to precisely what we need. Propel applies this to research workflows where the cost of undetected noise is highest — a silent broadcasting bug in a loss function doesn't crash, it produces subtly wrong training runs that waste compute.

Core Principles

Before the skills, gates, and agents — Propel is built on five non-negotiable principles that are injected into every session. These define the mindset, not the process.

Vision: Beyond Claude Code

While Propel is currently implemented as a Claude Code plugin, its core design — structured gates, investigation-first methodology, domain-specific auditors, and living documentation — is not tied to any particular agent platform. The pipeline encodes how researchers should interact with code-generating agents, regardless of the underlying model or tool.

The same constraint framework that prevents Claude Code from producing noisy averages can prevent any research code agent from doing the same. Propel's gates, questioners, and auditors represent a general protocol for human-AI collaboration in research engineering.

The goal is a portable specification: structured intake, grounding questioners, investigation scaffolds, paper-alignment auditing, and experiment registries that work across agent systems — Claude Code, Cursor, Copilot Workspace, Devin, or whatever comes next.

The Pipeline

Propel enforces five human-in-the-loop gates, two questioner checkpoints, and automatic auditor dispatch after every code change. The pipeline ensures investigation happens before implementation, design decisions are explicit, and silent bugs are caught before they reach training.

At each gate, the agent stops and asks structured questions that reveal design assumptions — never "shall I proceed?" but "should we [A] or [B]? A means [trade-off], B means [trade-off]." The questioners (Q0, Q1) ground the work in concrete reference implementations and specific implementation details before the agent acts.

Four Modes

Not every session needs the full pipeline. Choose a mode at the start of each session to filter which skills and gates are active. Each mode is a standalone tool — use Researcher just for literature, Debugger just for diagnosis, or Trainer just for launching runs.

See It In Action

Each mode shapes how the agent interacts with you. Watch simulated sessions for each mode.

When Would I Use This?

Three real-world starting points and how Propel guides you through each one.

Key Features

Skills

Propel includes 15+ skills organized by workflow phase. Each skill activates on specific triggers and guides the agent through a structured process.

Agents (Auditors)

Domain-specific auditors auto-dispatch after code changes. They run as subagents with separate context windows and read-only access.

Get Started

git clone https://github.com/KevinBian107/propel.git
cd propel && pip install -e .

cd /path/to/your/project
propel init

Then start Claude and run /intro to select a mode and set up your project. See the full Getting Started guide for a complete walkthrough.

What Only You Can Provide

Propel's constraints are necessary but not sufficient. The framework forces the agent to stop and ask structured questions at every phase transition, but the quality of the output depends entirely on what you bring to those checkpoints:

Acknowledgments

Propel combines ideas from multiple sources:

• obra/superpowers — Plugin architecture, discipline enforcement, verification gates, micro-task planning.
• scott-yj-yang/new-prompt — Session management CLI, auto-detection of project root, investigation artifact linking.
• Talmo's sleap-io — Investigation skill template with structured scratch/ patterns and living READMEs.
• Sionic AI's experiment registry — Retrospective skill and experiment learning workflow.
• brunoasm's claude skills — Think-deeply anti-sycophancy skill and PDF extraction.
• Weizhena's Deep-Research — Structured literature review with human-in-the-loop checkpoints.
• Context Engineering Template — Basic Claude Code usage patterns and context engineering principles.