Copilot Fluency Score
July 6, 2026 · View on GitHub
The Copilot Fluency Score maps your GitHub Copilot usage patterns from the last 30 days into a maturity model with 4 stages across 6 categories. The overall fluency stage is the median of the 6 category scores.
Stages
| Stage | Label | Description |
|---|---|---|
| 1 | Copilot Skeptic | Rarely uses Copilot or uses only basic features |
| 2 | Copilot Explorer | Exploring Copilot capabilities with occasional use |
| 3 | Copilot Collaborator | Regular, purposeful use across multiple features |
| 4 | Copilot Strategist | Strategic, advanced use leveraging the full Copilot ecosystem |
Categories
1. 💬 Prompt Engineering
Measures how you interact with Copilot through prompts, slash commands, and mode diversity.
| Stage | Criteria |
|---|---|
| 1 | Fewer than 5 total interactions |
| 2 | At least 5 total interactions (ask + edit + agent) |
| 3 | 30+ interactions and (2+ slash commands used or agent mode used) |
| 4 | 100+ interactions and agent mode used and (model switching or 3+ slash commands) |
Boosters (can raise the stage independently):
- Average 3+ exchanges per session → at least Stage 2
- Average 5+ exchanges per session → at least Stage 3
- Model switching across tiers → at least Stage 3
- Any Auto model usage → at least Stage 2; Auto model used in 50%+ of sessions → at least Stage 3; Auto model used in 80%+ of sessions (with 5+ sessions) → Stage 4
Recognised slash commands: /explain, /fix, /tests, /doc, /generate, /optimize, /new, /newNotebook, /search, /fixTestFailure, /setupTests
2. 📎 Context Engineering
Measures how you provide context to Copilot using explicit references.
| Stage | Criteria |
|---|---|
| 1 | No context references used |
| 2 | At least 1 context reference |
| 3 | 3+ different reference types and 10+ total references |
| 4 | 5+ different reference types and 30+ total references |
Tracked reference types: #file, #selection, #symbol, #codebase, @workspace, @terminal, @vscode, #clipboard, #changes, #problemsPanel, #outputPanel, #terminalLastCommand, #terminalSelection, image attachments (copilot.image), prompt files (promptFile), custom prompt commands (prompt)
Evidence: All tracked reference types are shown in the evidence panel when used (not just the basic ones). The panel also shows the total lines of code referenced through #file: range selections.
Boosters: Using image references (copilot.image) → at least Stage 3; using prompt files (promptFile) → at least Stage 3
Stage 3 hint behaviour: The "try specialized context variables" tip is dynamic — it only lists the specific variables the user hasn't tried yet. If the user has already used 2 or more of the specialized set (image attachments, prompt files, custom prompt commands, #changes, #problemsPanel, #outputPanel, #terminalLastCommand, #terminalSelection, #clipboard, @vscode), the hint is suppressed entirely.
3. 🤖 Agentic
Measures adoption of autonomous, multi-step agent mode workflows.
| Stage | Criteria |
|---|---|
| 1 | No agent-mode interactions |
| 2 | At least 1 agent-mode interaction |
| 3 | 10+ agent-mode interactions and 3+ unique intentional tools used |
| 4 | 50+ agent-mode interactions and 5+ unique intentional tools used |
Boosters:
- Multi-file edit sessions detected → at least Stage 2
- Average 3+ files per edit session → at least Stage 3
- 20+ multi-file edits with average 3+ files per session → Stage 4
- Sessions with 2+ child workspaces (multi-agent orchestration) → at least Stage 3; 3+ such sessions → at least Stage 4
Note about multi-agent orchestration: This signal is read from
~/.copilot/data.db(Copilot app only). It counts sessions where you were the parent of 2 or more child workspaces — covering both sessions where you manually launched child sessions and sessions where an agent autonomously spawned a subagent fleet. It is absent (and does not affect scoring) whendata.dbis not available.
Note: Only intentional tools count toward the unique tool thresholds — tools that Copilot calls automatically (file reads, searches, error lookups, confirmations, memory, etc.) are excluded. See Automatic vs. Intentional Tools below.
4. 🔧 Tool Usage
Measures breadth and depth of tool integration, including MCP servers.
| Stage | Criteria |
|---|---|
| 1 | No intentional tools used |
| 2 | At least 1 intentional tool used |
| 3 | 2+ advanced tools used, or @workspace agent sessions detected, or any MCP server usage |
| 4 | 2+ MCP servers used |
Recognised advanced tools: GitHub Pull Request, GitHub Repository, Run In Terminal, Edit Files, List Files
Note: Only intentional tools count toward Stage 2. Automatic tools are still shown in the tool-usage table with an
autobadge, but are not counted for scoring. See Automatic vs. Intentional Tools below.
Automatic vs. Intentional Tools
Copilot calls many tools on its own during agentic sessions to gather context — reading files, searching the codebase, checking errors, etc. These are called automatic tools and do not count toward fluency scoring because they do not reflect deliberate configuration choices by the user.
Automatic tools (excluded from fluency scoring):
- File operations:
read_file,list_dir,ls,view,find_files,glob,grep,grep_search,file_search,file_glob_search - Codebase search:
semantic_search,code_search,search_workspace_symbols,get_symbols_by_name - Project info:
get_errors,get_changed_files,read_project_structure,get_project_setup_info,get_vscode_api,get_doc_info - Terminal reads:
terminal_selection,terminal_last_command,get_terminal_output,await_terminal - Internal/session:
memory,detect_memories,tool_replay,vscode_get_confirmation*,ask_questions,switch_agent,bash
Intentional tools (count toward fluency scoring) include:
- Terminal execution:
run_in_terminal,run_build,run_task - File writing/editing:
edit_files,write_file,create_file,apply_patch,insert_edit_into_file,replace_string_in_file - Tests and runs:
runTests,run_notebook_cell,run_vscode_command,create_and_run_task - External integrations:
fetch_webpage,webfetch,websearch, MCP tools (all) - GitHub:
github_pull_request,github_repo - Browser:
open_integrated_browser,renderMermaidDiagram - Extensions and packages:
install_extension,install_python_packages
The full list of automatic tool IDs is maintained in src/automaticTools.json.
5. ⚙️ Customization
Measures how you tailor Copilot to your projects (custom instructions, model selection).
| Stage | Criteria |
|---|---|
| 1 | No repositories with customization (e.g. .github/copilot-instructions.md) |
| 2 | At least 1 repository with customization |
| 3 | 30%+ of repositories customized (minimum 2 repos) |
| 4 | 70%+ of repositories customized (minimum 3 repos) |
Boosters:
- Using 3+ different models → at least Stage 3; using 5+ models with 3+ customized repos → Stage 4
- Microsoft Foundry / local model usage → at least Stage 2
- Unknown provider model usage → at least Stage 2
6. 🔄 Workflow Integration
Measures how deeply Copilot is woven into your daily coding workflow.
| Stage | Criteria |
|---|---|
| 1 | Fewer than 3 sessions in the last 30 days |
| 2 | 3+ sessions |
| 3 | Using multiple modes (ask + agent) or 20+ explicit context references |
| 4 | 15+ sessions and multi-mode usage and 20+ context references |
Booster: 50%+ code-block apply rate → at least Stage 2
Overall Score Calculation
The overall fluency stage is the median of the 6 category scores:
- Sort all 6 category scores
- Take the average of the two middle values (since 6 is even)
- Round to the nearest integer
For example, if your category scores are [1, 2, 3, 3, 4, 4], the two middle values are 3 and 3, so the overall stage is 3 (Copilot Collaborator).
Data Source
Scores are calculated from Copilot Chat session log files stored locally on your machine. Some Copilot features (e.g., inline suggestion acceptance rates) are not captured in these logs and therefore not reflected in the score.
The dashboard updates every 5 minutes automatically. You can also refresh manually from the Fluency Score panel.