Copilot Playbook

When asked to generate a Copilot CLI playbook (a teaching guide for how the user drives the agent), apply this domain knowledge.

Copilot CLI Playbook — Domain Knowledge

Purpose

Produce a qualitative teaching guide — "how I drive the agent" — that someone else could read and adopt. It is the narrative sibling of the copilot-usage-report skill:

A playbook teaches a style, not a feature list. Every lesson must include real, copyable prompts drawn from actual session history — not invented examples.

Data Sources (use both)

1. The latest usage report (quantitative backing)

Read the newest Reports/<date>-copilot-usage-report.md first. Use it for:

• The volume line in the intro ("distilled from ~N sessions / M+ prompts").
• Which patterns are prominent — the report's verb-pattern table and "Key Behavioral Patterns" tell you which lessons matter most for this user (e.g. if create a plan… and implement… are the top two verbs, the plan→implement lesson leads).

Get-ChildItem Reports\*-copilot-usage-report.md | Sort-Object Name | Select-Object -Last 1

If no usage report exists, generate one first (see the copilot-usage-report skill) or derive the volume numbers from a session-count query.

2. Session history (the copyable prompts)

Mine session_store_sql for verbatim real prompts per behavioral pattern. These become the "Real prompts" bullets in each lesson.

IMPORTANT — query one pattern at a time. A single query that UNIONs many ILIKE patterns over a wide window times out. Run each pattern as its own small, time-filtered, LIMITed query, and ORDER BY RANDOM() for a representative sample:

-- Plan-first
SELECT substr(user_message,1,90) msg FROM turns
WHERE timestamp > now() - INTERVAL '30 days'
  AND user_message ILIKE 'create a plan%'
  AND length(user_message) BETWEEN 12 AND 90
ORDER BY RANDOM() LIMIT 6

Pattern filters (one query each):

• Plan → implement: ILIKE 'create a plan%'; separately ILIKE 'implement%' AND length < 40.
• Commit granularity: ILIKE '%commit%' AND (ILIKE '%per todo%' OR ILIKE '%per step%' OR ILIKE '%one commit%' OR ILIKE '%don''t commit%').
• Investigate before directing (questions): one query per opener — ILIKE 'does%', 'is %', 'what%', 'why%', 'how%'.
• Terse corrections / redirects: ILIKE 'no %', ILIKE 'don''t%', ILIKE 'instead%', ILIKE '%should be%', ILIKE '%not %use%'.
• PR lifecycle: ILIKE '%draft PR%', ILIKE '%PR description%', ILIKE '%resolve%comment%', ILIKE '%rebase conflict%'.
• Skills / memory: ILIKE '%skill%', ILIKE '/skill%', ILIKE '%remember%'.
• Verification: ILIKE '%check that%', ILIKE '%docs are up to date%', ILIKE '%what is remaining%', ILIKE '%dry-run%', ILIKE '%what are the local changes%'.

Prefer short prompts (length BETWEEN 12 AND 90) so the examples are clean and copyable. Dedupe near-identical results; quote verbatim (don't paraphrase).

3. Equipped & used plugins / MCPs (the toolset)

A playbook should show what the agent is equipped with and which of it actually gets reached for — the toolset is part of the working style.

Installed (the equipped set):

copilot plugin list
copilot plugin marketplace list
$mcpPath = Join-Path $HOME '.copilot\mcp-config.json'
if (Test-Path $mcpPath) {
    (Get-Content $mcpPath -Raw | ConvertFrom-Json).mcpServers.PSObject.Properties.Name
}

List installed plugins and registered marketplaces by name, not just as a count. Separate always-installed plugins from repository-local plugins loaded for a specific task.

Used (what actually gets invoked) — from session_store_sql's tool_requests:

-- Top tools; MCP tools are prefixed (github-mcp-server-, Playwright-, NuGet-, ...)
SELECT name, COUNT(*) c FROM tool_requests
WHERE session_id IN (SELECT id FROM sessions WHERE updated_at > now() - INTERVAL '30 days')
GROUP BY name ORDER BY c DESC LIMIT 40

Bucket the top-N by MCP prefix in-memory — a wide CASE-by-prefix aggregate can time out; pull the top list and sum prefixes such as github-mcp-server-*, Playwright-*, and NuGet-* yourself. The skill tool count is your skill-invocation total. Report the handful that dominate, not an exhaustive list.

Report Structure

Mirror Reports/2026-06-10-copilot-playbook.md:

1. Title + intro — # A Copilot CLI Playbook — How I Drive the Agent, then an italic line citing the volume and pointing at the usage reports for the underlying data. One sentence framing it as an adoptable style, not a feature list.
2. Lessons — one ## section per pattern, ----separated, each with:
   • Principle — a one-line imperative (the habit).
   • Why it works — 2-3 sentences on the payoff.
   • Real prompts: — 3-5 copyable bullets in backticks, pulled verbatim from history.
   • > Teaching point: — a blockquote with the transferable takeaway.
3. Quick-Start Checklist — a numbered list distilling every lesson into one actionable line each, for someone adopting the style cold.
4. Closing pointer — italic line: "Want the numbers behind these patterns? See Reports/<date>-copilot-usage-report.md."

Order the lessons by prominence for this user (from the usage report), but the durable core set is: Plan→Review→Implement, commit granularity, investigate-before-direct, terse corrections, short prompts, own-the-PR-lifecycle, skills+memory, equip-the-agent (plugins/MCPs), verify-don't-trust.

The Equip the Agent lesson is where plugins and MCP data lands: teach keeping a curated plugin set and enabling repository-specific tools only when needed. List the actual plugins and cite used MCP categories such as code search or browser automation as evidence of which tools earn their keep.

Writing Guidance

• Teach the style, not the tool. No feature dumps — each lesson is a habit with a rationale.
• Every lesson needs real prompts. If a pattern has no good real examples in history, drop the lesson rather than invent prompts.
• Keep it adoptable. Write for a reader who wants to work this way, not for the original author. Parameterize volume/patterns from the data — don't hardcode another user's numbers.
• Cross-reference, don't duplicate. The playbook cites the usage report for numbers; it doesn't reproduce the tables.

Output

Save to Reports/<YYYY-MM-DD>-copilot-playbook.md (date format YYYY-MM-DD).

Refreshing an Existing Playbook

When a prior *-copilot-playbook.md exists:

• Keep the durable lessons (the core set rarely changes).
• Refresh the prompts with newer verbatim examples and update the volume line from the latest usage report.
• Add lessons for newly-emerged patterns (e.g. a new ritual that now recurs) and drop any that no longer show up in history.
• Note at the top that it supersedes the prior dated playbook.