Copilot AI Credit Lab

A screenshot-by-screenshot practical guide

Created by Akash Dwivedi

Complete a real-looking checkout task in the browser, change six optimization settings, and compare the modeled AI credits. No installation, sign-in, shell, or AI subscription is required for this simulator.

Open the live lab   |   Open the CLI view

Screenshot convention: orange outlines mark the control to click. Screenshots are cropped from the deployed app, not mock-ups. All amounts are fixture estimates, not measured GitHub usage or promised savings.

Live workflow verified: 2026-09-23. Recommended setup: desktop browser at 100% zoom, approximately 1440 pixels wide.

Find the task you need

Follow one guided route in order, then apply the patch, compare, and export. Each task includes an actual screenshot plus a completion check.

Open a clean checkout demo

1. Open the live lab. Wait for the code editor to load. A first visit can take longer while the hosted app starts.
2. In YOUR SCENARIO, select Fix a checkout regression. If the selector is disabled, exit the existing tour with its top-right X first.
3. Click VS Code above the workspace. The selected surface stays selected when you start the tour.
4. Click Guided tour in the top-right page header. This resets the current demo and starts step 1 / 9.
5. Keep the default prompt and leave Assume warm instruction cache unchecked throughout the main walkthrough.

Save the baseline run

Capture the unoptimized request before changing any settings.

1. In the green tour panel above the editor, click Run baseline (outlined below).
2. Wait for the action to finish. The same button changes to Continue.
3. Check the result below, then click Continue once to open the next task.

Native editor control: Send prompt

Outside the tour, keep the prefilled prompt, then click the upward-arrow Send prompt button at the bottom-right of Copilot Chat. Hover to confirm its tooltip. Do not apply the patch yet.

Optional alternative. The tour button already performs this change.

Focus on the failing files

Keep the source and failing test; remove unrelated repository context.

1. In the green tour panel above the editor, click Attach task files (outlined below).
2. Wait for the action to finish. The same button changes to Continue.
3. Check the result below, then click Continue once to open the next task.

Native editor control: Attach task files

Click the paperclip at the bottom-left of the Chat composer, tooltip Attach task files. The simulator attaches the source and test automatically; there is no file chooser.

Optional alternative. The tour button already performs this change.

Keep only the relevant tools

Stop including schemas for tools this task does not need.

1. In the green tour panel above the editor, click Trim unrelated tools (outlined below).
2. Wait for the action to finish. The same button changes to Continue.
3. Check the result below, then click Continue once to open the next task.

Native editor control: Configure tools

Click the sliders icon beside the paperclip, tooltip Configure tools. In the dialog, uncheck the checkbox beside enterprise-integrations. Leave Core workspace tools enabled. Click Done.

Optional alternative. The tour button already performs this change.

Start a fresh conversation

Do not carry unrelated conversation history into a new task.

1. In the green tour panel above the editor, click Start new conversation (outlined below).
2. Wait for the action to finish. The same button changes to Continue.
3. Check the result below, then click Continue once to open the next task.

Native editor control: New chat

Click the + at the top-right of the Copilot Chat pane, tooltip New chat. Do not use the page-level Reset demo button.

Optional alternative. The tour button already performs this change.

Use lean project instructions

Retain essential engineering rules without loading a full handbook.

1. In the green tour panel above the editor, click Use lean instructions (outlined below).
2. Wait for the action to finish. The same button changes to Continue.
3. Check the result below, then click Continue once to open the next task.

Native editor control: instruction file

In the left Explorer, open .github > copilot-instructions.md. Click Use short, stable instructions below the editor. This fixture is read-only; use the button rather than trying to type into it.

Optional alternative. The tour button already performs this change.

Set medium reasoning effort

Match the modeled reasoning budget to this bounded regression.

1. In the green tour panel above the editor, click Set medium effort (outlined below).
2. Wait for the action to finish. The same button changes to Continue.
3. Check the result below, then click Continue once to open the next task.

Native editor control: Reasoning effort

Open the High effort dropdown in the bottom row of the Chat composer. Select Medium effort. Do not change the model dropdown yet.

Optional alternative. The tour button already performs this change.

Select Auto model routing

Demonstrate how a modeled change in pricing affects AIC.

1. In the green tour panel above the editor, click Select Auto (outlined below).
2. Wait for the action to finish. The same button changes to Continue.
3. Check the result below, then click Continue once to open the next task.

Native editor control: Chat model

Open the Claude Sonnet 4.6 dropdown above the Chat composer toolbar. Select Auto. Keep Medium effort selected.

Optional alternative. The tour button already performs this change.

Run the optimized request

1. Confirm all six levers are enabled and the inspector reads 3.61 AIC. Do not change the prompt or enable warm cache.
2. Click Run optimized task in the green panel. Wait for the simulated response and patch preview.
3. The primary button changes to Continue. Click it once.

Native alternative

Click the Chat composer upward-arrow Send prompt button, or the inspector's Run comparison button. Use only one, once. Each completed request adds a history row.

Finish without losing your work

1. At Take the habit back to your workspace., click Finish tour.
2. Do not click Guided tour again: it starts over and clears the saved runs.
3. Continue to Apply the patch and check it, then export your run report before leaving the tab.

Choose the terminal route

1. Open the CLI link, or click Copilot CLI above the workspace.
2. Select Fix a checkout regression in YOUR SCENARIO.
3. For the guided route, click Guided tour in the top-right header. Confirm 1 / 10.
4. Follow the next ten tasks in order. Use the tour action buttons for the exact published totals and two-run comparison.

Press Enter once to submit. Wait until the input is enabled again before the next command. These commands belong inside this browser demo, not PowerShell, Azure CLI, or gh copilot.

Save the baseline run

1. In the green tour panel, click Run baseline.
2. Wait for Continue to become enabled. Check the result, then click Continue once.

Manual CLI alternative

Outside the tour, click the clipboard icon beside the CLI input (Insert scenario prompt), then press Enter. Wait until the input is enabled again.

The baseline includes broad files, 28 tool schemas, prior conversation history, long instructions, high reasoning effort, and manual model selection.

Focus on the failing files

1. In the green tour panel, click Attach task files.
2. Wait for Continue to become enabled. Check the result, then click Continue once.

Manual CLI alternative

In the CLI input, type the following and press Enter. This both enables focused context and runs the fixture, so the manual route adds a run here.

@src/cart.ts @src/cart.test.ts Fix the discount regression. Keep the public API unchanged.

Keep the tour prompt unchanged to reproduce the exact AIC values. A custom CLI prompt has a different estimated size and may change the last decimal.

Keep only the relevant tools

1. In the green tour panel, click Trim unrelated tools.
2. Wait for Continue to become enabled. Check the result, then click Continue once.

Manual CLI alternative

Type this command into the CLI input and press Enter. The terminal confirms that three core schemas remain.

/mcp disable enterprise-integrations

enterprise-integrations is fictional. The modeled saving assumes eager schema loading; deferred loading in real clients can make this saving smaller or zero.

Start a fresh conversation

1. In the green tour panel, click Start new conversation.
2. Wait for Continue to become enabled. Check the result, then click Continue once.

Manual CLI alternative

Type the following and press Enter. The terminal clears and its header reports History: new session.

/clear

For ongoing work, /compact retains a 1,200-token summary. It is not equivalent to /clear and produces a higher modeled total. The eraser icon only clears the displayed terminal entries.

Use lean project instructions

1. In the green tour panel, click Use lean instructions.
2. Wait for Continue to become enabled. Check the result, then click Continue once.

Manual CLI alternative

Enter /instructions. In the Project instructions dialog, check Use short, stable project instructions. Close it with the top-right X. Downloading the instructions alone does not enable this lever.

Leave the warm-cache checkbox off for this walkthrough. Shorter instructions do not guarantee a provider cache hit.

Set medium reasoning effort

1. In the green tour panel, click Set medium effort.
2. Wait for Continue to become enabled. Check the result, then click Continue once.

Manual CLI alternative

Enter /model. In the picker above the CLI input, open Effort, select Medium, then click Done. Keep Model unchanged for this task.

The response is a fixed reference answer. This does not prove that lower reasoning effort preserves quality on arbitrary real tasks.

Select Auto model routing

1. In the green tour panel, click Select Auto.
2. Wait for Continue to become enabled. Check the result, then click Continue once.

Manual CLI alternative

Enter the following command. The terminal confirms Model set to auto. The footer shows Auto and medium.

/model auto

The simulator assumes a 10% lower blended rate for Auto. This is not a guaranteed discount or a statement about the model real Auto will select.

Run HydraFusion

1. At Orchestrate the task with HydraFusion., click Run HydraFusion in the green panel.
2. Wait for the workflow result. Do not click again while it is running.
3. Verify the result below, then click Continue.

Manual alternative

Enter /hydrafusion run in the input and press Enter. After completion, enter /hydrafusion status. The terminal should report Latest HydraFusion workflow: completed.

Recognize the saved optimized run

1. At Same task. Now run it lean., look at the primary tour button.
2. After HydraFusion it already says Continue. Click Continue once. You do not need to find or press Run optimized task.
3. If you skipped HydraFusion and genuinely see Run optimized task, click it, wait for completion, then click Continue.

Finish the CLI tour

1. Click Finish tour in the green panel.
2. Click Run history in the top page navigation. Confirm baseline and optimized rows.
3. Click Simulator to return. The next section applies and checks the reference patch.

Apply the patch and check it

1. From Simulator, click VS Code above the workspace. A response from the CLI is shared with the editor.
2. In Copilot Chat, click Apply reference patch. If it says Reference patch applied, skip this click.
3. Click the small triangle at the right edge of the editor tab bar, accessible name Run reference checks, tooltip Check reference fixture. This is not the Send prompt arrow.
4. Read the OUTPUT panel underneath the editor.

Read the saved results correctly

1. Click Run history in the page header.
2. Read Levers, Baseline AIC, Configured AIC, and AIC reduction for each row.
3. The newest run is at the top. The exact guided route gives two rows; manual prompts or repeated runs add more.
4. Click Simulator to return. Click Comparison in the inspector for the current unsaved configuration.

Saved versus current

Changing switches updates the inspector immediately but does not change old rows. Run comparison or Send prompt to save a new row.

For the checkout route, the headline is approximately 85% fewer modeled AIC, with the same supplied reference answer.

The optimized row compares its own prompt against all levers off. It is not a direct subtraction from every other row.

Quality is a separate check. Equal answers are built into this fixture. For real work, measure actual usage and independently validate output quality.

Download the run report

1. Complete at least one run. Open Run history.
2. Click Export report at the upper-right of the results table.
3. Open your browser Downloads list. Look for copilot-ai-credit-lab-report.json. It is JSON data, not this PDF.
4. You can also hover over the download-arrow icon near the page heading and click Export run report. Both controls export the same report.

Disabled button? No completed run exists yet. Run the baseline and wait for its simulated response, then retry the export. If the browser blocks a download, allow the download for this site.

Send a reproducible configuration

1. Click Simulator and leave the desired scenario, surface, and lever settings selected.
2. In the upper-right page action icons, click the connected-nodes icon, tooltip Share configuration.
3. In Share this configuration, click Copy link. Expect Link copied.
4. If clipboard access is blocked, click the Configuration URL field, press Ctrl+C on Windows, and paste it into Teams.
5. Ask the recipient to open the link and run the request. Attach your exported JSON separately when they need your saved run evidence.

Fix an API retry boundary

1. Export previous runs first. Exit any open guided tour with its X.
2. Select Fix an API retry boundary under YOUR SCENARIO, then click VS Code. Click Task brief to inspect the problem and close it with X.
3. For the full before/after exercise, click Guided tour and repeat the nine VS Code tasks. Scenario selection alone does not clear previous levers or history.
4. For a quick fix-only exercise instead: turn all six switches on (or click Enable all if shown), keep warm cache off, and click Send prompt.
5. When the response appears, click Apply reference patch, then the editor-tab Run reference checks triangle.

Why the fix works: The inclusive loop runs attempts 0, 1, 2, and 3. The reference change uses < instead of <= to make exactly three attempts while preserving the final error.

CLI variant: select this scenario, switch to Copilot CLI, and enter /hydrafusion run. It automatically uses src/retry.ts and src/retry.test.ts.

Cover a shipping edge case

1. Export previous runs first. Exit any open guided tour with its X.
2. Select Cover a shipping edge case under YOUR SCENARIO, then click VS Code. Click Task brief to inspect the problem and close it with X.
3. For the full before/after exercise, click Guided tour and repeat the nine VS Code tasks. Scenario selection alone does not clear previous levers or history.
4. For a quick fix-only exercise instead: turn all six switches on (or click Enable all if shown), keep warm cache off, and click Send prompt.
5. When the response appears, click Apply reference patch, then the editor-tab Run reference checks triangle.

Why the fix works: The reference change uses >= instead of > so exactly $50 qualifies for free shipping. The boundary tests are already supplied in this demo; you do not need to author them.

CLI variant: select this scenario, switch to Copilot CLI, and enter /hydrafusion run. It automatically uses src/shipping.ts and src/shipping.test.ts.

What the simulator counts

Read Breakdown in the inspector

System: fixed system-instruction fixture.

Instructions: long or lean project rules.

Tools: available tool schemas, not the number of tools actually executed.

History: old turns, an empty conversation, or a retained summary.

Files: broad retrieval or focused task files.

Your prompt: an estimate from entered character count.

Output + reasoning: fixed answer plus illustrative reasoning budget.

Pricing assumptions

The model converts estimated input and output usage to USD using its configured Sonnet 4.6 rates, then uses 1 AIC = $0.01 USD. Auto applies an illustrative 10% lower blended rate. These are simulator assumptions, not current-plan billing advice.

Optional cache experiment

1. Enable Keep instructions lean first. Otherwise the cache checkbox is disabled.
2. Check Assume warm instruction cache below the inspector. Observe the changed estimate, then uncheck it to restore the guide's values.
3. Click View assumptions below Run comparison for the complete methodology.

Get unstuck without restarting blindly

Only when you need a complete restart

1. Export existing runs first.
2. Click the circular-arrow page icon, tooltip Reset demo.
3. Read the dialog and click Reset demo again to confirm, or Cancel to keep your work.

Reloading also loses in-memory runs. The URL can restore configuration, not results. Real Copilot versions, plans, available models, tools, and organization policy may differ from this replica.