The Short Leash AI Coding Method for High-Quality Software

The Short Leash Method: Prioritizing Quality Over Autonomy

The Short Leash method is a disciplined approach to AI-assisted software development that rejects the trend of fully autonomous "vibe coding" in favor of strict human oversight. It is specifically designed for expert developers working on security-critical systems where software quality and a deep mental model of the codebase are non-negotiable.

While many current AI workflows emphasize removing the human from the loop via orchestrators and parallel agents, the Short Leash method argues that such approaches lead to "slop"—inefficient, ugly, or subtly broken code that the developer cannot effectively debug because they have lost their understanding of the system.

Core Principles of the Short Leash Workflow

To prevent AI agents from "going off the rails" and to ensure the resulting code beats the quality of raw output from frontier models like Fable 5, the Short Leash method implements the following constraints:

1. Strict Human-in-the-Loop Execution

No "YOLO" Mode: Developers must never use "dangerously skip permissions" settings. Every action the AI takes must be explicitly permitted.
Active Analysis: The developer must analyze the diff of every proposed change in the permissions prompt before granting access.
Constant Engagement: The AI does not work in the background while the developer is distracted. The developer remains present to intervene the moment the AI deviates from the intended path.
Frequent Denials: Permissions should be denied immediately if the AI proposes a change that is undesirable or incorrect.

2. Structured Planning and Versioning

Formal Planning Phase: Every task begins with a research and planning phase to formulate a concrete strategy.
Task Tracking: Large tasks are broken into smaller, trackable steps to maintain focus.
Frequent Commits: Commits are made at the end of every subtask. This prevents the AI from accidentally deleting previously completed work, a failure mode observed even in high-end models like Opus.

Rigorous AI Review Process

High-quality software requires a dual-layer review process. The Short Leash method posits that a PR reviewed by both a human and an AI contains fewer mistakes than one reviewed by either alone.

The AI as a Linter

AI should be used to review every PR to catch common mistakes and syntax errors quickly. For this to be effective, the AI must have full context, including the issue description, the PR description, the codebase, and the specific changes.

Human Accountability and Disclosure

AI Disclosure: Every PR must include an "AI Disclosure" heading specifying the exact models used. This informs maintainers of the tools used and signals transparency.
Mandatory Self-Review: If AI was used to generate the code, the author must review their own PR line-by-line as if they were reviewing a stranger's work. This ensures the human author fully understands the submitted code and maintains their mental model of the codebase.

Community Perspectives and Counterpoints

While the Short Leash method emphasizes rigor, developers in the community have raised several counter-arguments regarding efficiency and the nature of AI capability:

The "Crutch" Argument

Some developers argue that micromanaging a frontier model is a waste of time. One contributor noted that providing more nuance and having deeper design discussions with models like Opus can lead to better architectural suggestions (e.g., suggesting a MILP library over a greedy solver) without needing a "short leash" for every line of code.

The "Junior Engineer" Analogy

Another perspective suggests treating the AI as a junior-to-mid-level engineer. In this model, the AI is given a sandbox (isolated VM) and "YOLO" permissions to work on a feature up to a PR-worthy point, at which point the human performs a rigorous diff review. This avoids the mental exhaustion of micromanagement while maintaining final quality control.

The Mental Model Debate

There is significant disagreement on whether a human still needs to understand the codebase. While the Short Leash method views this as essential, some argue that we are moving toward a future where humans no longer need to understand the underlying code, letting the AI drive the system entirely.

Productivity Trade-offs

Some users report that the Short Leash method provides little to no speedup, suggesting that productivity comes from outsourcing as much work as possible to the model in a sandbox and minimizing the effort required to review the final output.