/immersive

A slash prompt that turns a vague "I wish Copilot would…" into a concrete, scaffolded customization. It interviews you about a desired improvement, picks the right Copilot primitive (instruction · prompt · skill · agent · hook · tool), justifies the choice, and writes the files under .github/.

How it works

Type /immersive <describe the improvement> in Copilot Chat. With no trailing text it asks you to describe the improvement and stops. Otherwise it follows a fixed workflow:

1. Restate the improvement in one sentence.
2. Produce a Diagnose block — who triggers it, how often, knowledge vs. capability, persona needed, decision-tree walk, recommendation, trade-offs.
3. Confirm if the recommendation is non-obvious.
4. Scaffold the file(s) under .github/.
5. Run get_errors on the new files and summarize: workspace links, how to activate, one smoke test, one promotion follow-up.

The full prompt body lives in immersive.prompt.md; the decision rules live in the companion immersive chatmode.

Examples

Example 1 — Gang of Four design patterns

"Copilot keeps suggesting design patterns but never names them, and picks the wrong one for our codebase."

/immersive diagnosed this as domain knowledge + always-on rules:

• 23 GoF patterns + a decision tree → a skill that loads on demand: .github/skills/gof-design-patterns/SKILL.md
• Naming convention + trade-offs are mandatory → an instructions file that's always in context: .github/instructions/design-patterns.instructions.md

Why two primitives: the catalog is too big to keep loaded all the time (skill), but the rules about how to talk about patterns are short and must apply everywhere (instructions).

Example 2 — format-on-edit

"Every time Copilot generates code, the formatter doesn't run and PRs end up full of whitespace diffs."

/immersive diagnosed this as a post-event capability — nothing the model itself should decide, just something that must happen after every edit. Recommendation: a hook.

• .github/hooks/format-on-edit.json wires the post-edit event to a polyglot dispatcher.
• .github/hooks/scripts/format-on-edit.ps1 switches by file extension to prettier / black / gofmt / rustfmt / dotnet format.

Why a hook (not a tool): the agent shouldn't have to decide to format — the runtime fires it. Non-blocking, output to log, no return value to the model. That's the Hook ≠ Tool distinction /immersive enforces.

When to reach for /immersive

• You catch yourself re-typing the same context every chat → probably a prompt or instructions file.
• A teammate asks "which pattern fits?" and you have a real answer → a skill.
• Something needs to happen after every Copilot edit → a hook.
• The model needs a deterministic capability it doesn't have → an MCP tool.
• A workflow has a fixed persona and tool allowlist → an agent.

/immersive walks you through picking among these instead of guessing.