Passive Prompt Coaching for Copilot CLI

The issue describes a new feature for passive prompt coaching in Copilot CLI. Start by exploring the existing CLI codebase to understand how prompts are processed and sessions are tracked. Look for entry points related to user interaction and AI response handling. The mockup plugin linked in the issue may provide initial implementation ideas. 'Done' means a coaching system that observes patterns and offers non-disruptive recommendations.

Describe the feature or problem you'd like to solve

Prompt-coaching provides passive coaching in real time by observing prompting patterns across an entire session and offering contextual recommendations without disrupting the developer's workflow

Proposed solution

AI harnesses have become incredibly capable. It is remarkable to think that developers no longer need to be nearly as explicit, verbose, or specific in their instructions as they did just over a year ago.
However, we still need to be mindful that complacency can creep in over time, particularly as we become more accustomed to relying on AI.
Relevant research:

• Thinking Fast, Slow, and Artificially
• The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers

Developers can also be wasteful with their AI token spend when they consistently fall into poor prompting habits, such as “prompt and pray”, failing to define success criteria, or providing ambiguous instructions that leave room for unintended or misleading results.
The challenge is that developers rarely stop mid-flow to reflect on the quality of their prompting.
This feature would provide passive, real-time coaching by observing prompting patterns across an entire Copilot CLI session and offering contextual recommendations without disrupting the developer's workflow.
Importantly, the feature would not block a task or automatically rewrite prompts. Instead, it would identify behavioural anti-patterns over the course of a session and coach developers towards more effective AI collaboration.
For example, it could identify recurring patterns such as:

• repeatedly issuing corrective follow-up prompts because the original request lacked clear success criteria;
• providing ambiguous instructions that repeatedly produce unintended results;
• repeatedly re-explaining context that could have been established more effectively;
• using a “prompt and pray” approach rather than giving the agent clear constraints or validation criteria.
  The goal is not to judge individual prompts, but to help developers recognise patterns in how they collaborate with AI over time, improving prompting habits while preserving the natural flow of development.

Example prompts or workflows

Potential examples of helpful advice

To activate:

/coaching on

_User Prompt: Analyse this dataset_
_Coach: Your prompt requests analysis but doesn't specify the outcome. Consider adding desired metrics, trends, anomalies, or business questions._

Coach: You've submitted 5 successive prompts that rephrase the request without adding context. Consider providing error details, sample output, or acceptance criteria.

Coach: This prompt contains multiple objectives. Breaking it into smaller tasks may improve reliability and reduce token usage.

Coach: Large amounts of unrelated context were included. Omitting irrelevant information may reduce cost without impacting answer quality.

Coach: You've accepted several generated solutions without modification. Consider validating assumptions before implementation_

Additional context

• Hackathon Project
• Mockup-plugin