The harness beats the tricks: GitHub's AI productivity recipe

GitHub's post is a manifesto against complexity in AI tooling, written by someone who works with AI daily and has concluded that less is more. The author observes that the biggest productivity gains do not come from installing new tools, configuring obscure features, or tricking agents with clever prompts; instead, they come from understanding and using the harness—the core interface—well.

The recommended workflow begins with choosing a tool. Within the GitHub Copilot family, options include the CLI, the new GitHub Copilot app, VS Code, Visual Studio, and JetBrains. The author notes that these experiences are increasingly centralized on the same underlying harness, so learning it once applies everywhere. For beginners, the CLI is recommended because it is text-only and offers the most direct interaction with the agent.

Step two is to enable YOLO mode (also called "Allow All"), which lets the agent execute any command without asking permission each time. The author explains that agents need autonomy to drive productivity; constant approval is both a miserable user experience and counterproductive, training users to stop reading what they are asked to approve. However, the author emphasizes safety: YOLO mode should only be used in sandboxes such as GitHub Codespaces or development containers, never on a local machine or organizational systems where mistakes could expose private data.

Step three is to start with a prototype. The author illustrates this with a date picker web component example. Instead of diving into implementation, the author asks the AI to generate 20 mockups in a single HTML file so they can be compared. One mock—which displays a year view that can zoom into months and then days—sparked an insight the author would not have considered upfront. The author generalizes that humans process sensory-rich models like images and layouts faster than dense text, so low-effort prototypes early on help make complex concepts immediately intuitive. This applies even to non-visual tasks like API design, where the author asks the AI to create a visual mockup with five options for implementing a new API endpoint, rendered as Mermaid diagrams.

Step four is methodical planning. The author switches to plan mode in GitHub Copilot with a prompt like "/plan Build a date picker web component. I want the user to be able to zoom in and out of years, months, and days." The model asks clarifying questions—Can the start and end date be the same? Are partial selections valid? Should users be able to clear the date?—that help surface edge cases and dependencies. The author notes that the point is not to accept every suggestion but to deeply engage with the problem and guide the model, using human expertise. The author also mentions the "grill-me" skill from Matt Pocock, which makes the planning phase more aggressive by asking even more questions and edge cases.

Step five is implementation with Autopilot. Once planning is done, the model automatically switches to Autopilot, which is a built-in loop that forces the model to continue working and verify that every item in the plan has been completed. GitHub Copilot automatically orchestrates this phase: it uses the "Explore" subagent with a small model to read files, and the "General Purpose" subagent with a larger model for complex actions. All of this happens out of the box without manual configuration.

Step six is human review and iteration. The author acknowledges that the AI output will likely not be exactly what was wanted, and iteration is where human taste decides the final quality. In the date picker example, the AI-generated component had inconsistent animations, unreadable text on hover due to color contrast, unnecessary labels, and a design that looked "too much like it was created by AI—because it was!" The author then iterates, adding guidance like using a CSS framework called Postrboard and asking for fixes to specific behaviors (e.g., clicking "Today" should take you to day view, not stay in month view). The author emphasizes not settling for "good enough" and being ruthless about quality.

Step seven is a rubber-duck review, where the author asks GitHub Copilot to perform a rubber-duck code review of the final result. The author does not detail what comes next but indicates this is the final checkpoint.

Throughout, the author recommends using a medium-sized model such as GPT 5.6 Terra or Claude Sonnet with medium reasoning for most work, and sticking with the same model during a feature or bug fix to maximize prompt caching savings and reduce token costs. The overarching message is that the harness—mastered, used consistently, and paired with human judgment—is all you need; exotic prompts and skills are mostly gimmicks.

The post emerges from a frustration with AI hype—the endless stream of "one weird trick" posts and tool announcements that promise mastery without substance. What GitHub is articulating is that productivity with AI agents comes not from novelty but from depth: learning one interface thoroughly, understanding its affordances, and applying a consistent methodology.

The seven-step workflow itself is grounded in familiar software practices: prototype before committing, plan edge cases, implement systematically, review rigorously, and iterate. What is new is that AI makes prototyping almost free—you can generate 20 design mockups in seconds—and planning can be crowdsourced to the model itself, which helps surface questions a human planner might miss. The author treats the model as a thought partner, not an oracle: the model asks questions, the human decides what matters, and the model implements. This division of labor—human judgment on quality and intent, AI on execution and ideation—appears to be where the real gain lies.

A second theme is the tension between autonomy and safety. YOLO mode (unrestricted command execution) is essential for productivity, but only safe in a sandbox. This suggests that the future of AI at work will depend less on perfect prompt engineering and more on infrastructure that permits trust: development containers, ephemeral environments, and clear boundaries between AI workspaces and production systems.

No comments yet. Be the first to share your thoughts!