VS Code + Copilot — Setup Guide
For Topher's SyF second-brain work. Builds on the
copilot-ide-research.md
decision doc (2026-04-28, IntelliJ-vs-VS-Code). Now that VS Code is the path,
this is how to actually set it up. Current through 2026-05-18.
What's changed since the April baseline
The 2026-04-28 research is still mostly accurate, but five updates matter:
| Date | What | Why it matters to you |
|---|---|---|
2026-05-17 | GPT-5.3-Codex became the base model for Copilot Business + Enterprise | 25% faster on agentic tasks than 5.2-Codex. SWE-Bench Pro + Terminal-Bench new SOTA. You get this automatically — no setting to flip. |
2026-04-13 | US + EU data residency + FedRAMP-authorized models in Copilot | Inference + data stays in the chosen geography. 10% premium-request multiplier when enabled. The single most important fact for SyF compliance review. |
2026-03-25 | GitHub Copilot for Jira public-preview enhancements | Copilot can read Jira issues + Confluence pages as context. PMs feed context directly from Jira. MCP Atlassian connector enables this. |
2026-02-XX | Atlassian Rovo MCP server GA | 72+ tools across Jira, Confluence, Compass. OAuth-scoped to your Atlassian permissions. This is the single integration that unlocks your day-job context. |
2025-11-01 | Copilot Knowledge Bases retired → replaced by Copilot Spaces | The Apr 28 research recommended Enterprise "for Knowledge Bases." That advice is now outdated. Spaces are better, available to anyone with a Copilot seat (even Free), and accept free-text content — transcripts, specs, vision drafts — not just markdown repos. |
Biggest single shift: the April baseline said "push for Copilot Enterprise to get Knowledge Bases." That feature was retired Nov 1, 2025 and replaced by Copilot Spaces — which are available on any paid seat (Business is fine) and support free-text content (specs, transcripts, vision drafts) which Knowledge Bases didn't. Business is now the right tier for you, not Enterprise. Save SyF the $20/seat/month delta.
The mental model — five customization surfaces
Once Copilot is installed, you have five places to "configure" it. Each does something different. Most setup guides conflate them, which is why most people use 10% of what's possible.
| Surface | Where | What it's for | Applies when | Example for you | Caveat |
|---|---|---|---|---|---|
| Custom Instructions | .github/copilot-instructions.md | Always-on context: who you are, what the project is, how you want Copilot to talk to you. | Every chat / inline / edit / agent interaction in this workspace. | SyF tone, Jira field conventions, "use Gherkin for AC", "never include named-employee coaching detail in any output." | Keep it short. <100 lines. Long instructions get partially ignored. |
| Path-Scoped Instructions | .github/instructions/*.instructions.md | Instructions that only apply to files matching a glob pattern (set in frontmatter). | When the active file matches the glob. | A user-story.instructions.md that applies to stories/*.md — forces Gherkin AC format. A vision-deck.instructions.md that applies to decks/*.md — enforces the "merchant pain → SyF answer → unlock" arc. | Newer feature (Q1 2026). Public docs are still thin. Worth piloting. |
| Prompt Files | .github/prompts/*.prompt.md | Reusable slash-commands. Type / in chat → pick one → it runs the pre-written prompt. | On-demand. Best for outputs you produce repeatedly. | /bijayta-monthly → builds the Bijayta deck outline from this month's sprint data. /po-coaching → drafts a quarterly coaching doc for one of your 4 POs. | Still public preview. Files live in workspace OR user profile. |
| Copilot Spaces | github.com/copilot/spaces (cloud) | Grouped context — code + free text + images + PR/issues — that Copilot grounds answers in. | When you ask in this Space's chat. Sharable with the team. | "Synchrony Pay Vision" Space — drop in the vision deck draft, merchant-routing wiki, Robin/Prem chat summaries. Ask Copilot to refine the pitch grounded in all of it. | Org-managed Spaces require admin promotion. Personal Spaces work today on any paid seat. |
| MCP Servers | .vscode/mcp.json (workspace) or user profile | Plug external tools into Agent mode. Jira, Confluence, GitHub, custom internal APIs. | Agent mode ONLY. MCP tools are invisible in Ask/Edit mode. | Atlassian Rovo MCP → ask Copilot "what are this sprint's stuck Jira tickets and why?" — it queries Jira live, summarizes, suggests next actions. | Requires VS Code 1.99+. Auth via OAuth (preferred) or PAT. Servers run locally as processes — check with IT before adding. |
Chat modes — Ask vs Edit vs Agent
These are the three things you can do in the Chat panel. Pick deliberately.
💬 Ask
Conversational Q&A. No file edits, no tool calls. Just answers.
You'll use it for: Brainstorming. "How should I frame the Synchrony Pay pitch for Mike S?"
✏️ Edit
Direct file edits with a preview-then-accept gate. You name the file scope.
You'll use it for: Refining a deck outline doc in place. "Tighten section 3 — too verbose."
🤖 Agent
Multi-step autonomous workflow. Can run commands, edit files, call MCP tools. Asks for confirmation before risky actions.
You'll use it for: Anything multi-step. "Pull this Confluence page, extract the routing logic, generate a SharePoint-friendly version, save it as routing-faq.md."
Default starting position: Ask mode for thinking, Edit mode for writing, Agent mode for anything that touches more than one file or calls an external tool. Always commit to git before kicking off an Agent run on important files.
Model picker — which one when
As of 2026-05-17, the Copilot default model is GPT-5.3-Codex for Business + Enterprise. You can override per-chat via the model picker at the bottom of the chat panel.
| Model | Premium multiplier | Best for | PM verdict |
|---|---|---|---|
| GPT-5.3-Codex (default) | 1× | Almost everything. Agentic workflows, multi-step reasoning, deck outlining, AC writing, story drafting. 25% faster than the predecessor. | Just leave it on default. Don't overthink this. |
| Claude Sonnet 4.6 | ~1× | Long-document reasoning, refactoring a verbose draft, "explain the tradeoffs" thinking work. | Switch to it when you're iterating on the Synchrony Pay vision deck and need the model to hold the whole document in its head. |
| Claude Opus 4.6 | Higher (premium) | The deepest reasoning tasks. Political pitch strategy, "how do I get Mike S to greenlight this" framing. | Use sparingly. Burns premium requests fast. Save for the moments that matter. |
| Gemini 3.1 Pro | ~1× | Multimodal — screenshots of merchant routing diagrams, deck images, dashboards. | Default to this when you're asking Copilot to react to an image. |
| Auto | Varies | Letting Copilot pick. It will route lightweight asks to cheaper models, hard asks to bigger ones. | Reasonable default if you don't want to think about it. Topher's recommendation: pick deliberately for important tasks; Auto for casual asks. |
The 80/20 rule: Default GPT-5.3-Codex covers ~80% of your work. Switch to Claude Sonnet 4.6 when you're iterating on the Synchrony Pay vision deck and the model needs to hold the whole document in its head. Don't fiddle with this otherwise.
Setup checklist
Pre-flight (do these once)
- Confirm SyF IT has approved BOTH github.copilot AND github.copilot-chat extensions
- Sign in: hover the Copilot icon in the Status Bar → Use AI Features → GitHub sign-in
- Check seat type: profile menu → Your GitHub Copilot. Confirm Business or Enterprise (not Pro/Free).
- Update VS Code to ≥1.99 (Agent mode + MCP require it)
- Ask SyF IT: is data-residency enabled on the Copilot policy? If you're in a regulated lane, push for US-residency.
Workspace setup (per project)
- Create a folder for your SyF second-brain work (e.g. ~/projects/syf-secondbrain). Open it in VS Code.
- Run /init in Copilot Chat — it scans the folder and drafts a copilot-instructions.md for you. Review + edit.
- Create .github/prompts/ — drop in the 2-3 prompt files for your highest-frequency outputs (Bijayta deck, PO coaching, AC review).
- Create .vscode/mcp.json — register the Atlassian MCP if approved. Otherwise wait.
- Commit this folder to a personal repo (private). Your customizations travel with you.
First real workflow (pilot)
- Pick ONE recurring output to automate: the Bijayta monthly deck is the highest-ROI candidate (4hr → 1hr).
- In Agent mode, paste last month's deck as context. Ask: "extract the structure, give me a prompt file that would generate next month's draft from a list of sprint summaries."
- Iterate on the prompt file until output is 80% there. Commit it.
- Next month: run the prompt, paste sprint data, get a draft in <10 min instead of 4 hours.
Compliance posture
- Never paste customer PII, account numbers, or merchant contract terms into Copilot — same M4 rules apply.
- Named-employee coaching detail (e.g. Rachel/Patrick candor) stays on Aria. Sanitize before Copilot sees it.
- Strategic frameworks, templates, sprint data, sanitized AC → all safe.
- Audit log gap: Copilot does NOT retain prompt/response content (privacy by design). Surface this to security if asked.
Starter .github/copilot-instructions.md
Drop this into your SyF second-brain workspace as a starting point. Tune over the first week — when Copilot misbehaves, the fix is usually a one-line clarification here.
# SyF Second-Brain — Copilot Custom Instructions ## Who I am Senior Train Leader at Synchrony Financial. Run Emerging Payments (dual card + cobrand) and lead vision on the EHS / UniFi train. 4 PO direct reports + 1 TL direct report. I am NOT a developer. I work in PowerPoint, Confluence, Jira, and prose. ## What I'm doing here This workspace is my personal second brain — sanitized SyF context, templates, prompt files, and drafting space. Everything here is Tier 1 (personal device, never shared). Outputs from here flow into OneNote (Tier 2), Confluence (Tier 3), or SharePoint (Tier 4). ## How to talk to me - Direct. Skip the throat-clearing. No "great question." - Bullets and tables over prose when comparing options. - When I ask for a deck, give me an outline I can paste into PowerPoint — not full slides. - When I ask for a user story, default to: As a / I want / So that, then Gherkin AC. - Always ask one clarifying question if my prompt is ambiguous. Don't guess. ## What to avoid - Never include named-employee coaching detail in outputs intended for company systems. - Never invent metrics. If I haven't given you the number, say "[METRIC TBD]". - Don't suggest tools that aren't on the SyF approved list — Copilot, ROVO, M365 Copilot Chat, SYF GPT only. ## Vocabulary I use - "The train" = Emerging Payments train (Card Access, Tokenization, DWAM, Provisioning Ops) - "The other train" = EHS / UniFi (Christina Berry leads ops, Nicole V is PM) - "Bijayta deck" = monthly leadership status, ~22 trains, ROI/sprint/analytics/blockers - "Synchrony Pay" = the rebrand-UniFi-to-multichannel-checkout moonshot - "Mike S" = senior exec over EP + AI; trusts me; informal channel today
Starter prompt file — Bijayta monthly deck
Save as .github/prompts/bijayta-monthly.prompt.md. Invoke in chat by typing
/bijayta-monthly (in the prompt-name dropdown). Goal: replace your 4-hour
monthly build with a <1-hour one.
--- mode: agent description: Draft the Bijayta monthly accountability deck outline --- You are helping me draft my monthly leadership accountability deck for Bijayta B (my skip at SyF). The deck covers ~22 trains and is presented monthly. ## Standard structure (one slide per train, 6 trains in my direct scope) 1. Train name + leader 2. ROI / value delivered this period (1-2 bullets) 3. Sprint status — on track / at risk / blocked 4. Analytics — key metric movement (or "[METRIC TBD]" if I haven't given you data) 5. Roadblocks needing skip-level support (be specific — name the blocker, not a vague "alignment") ## My ask Take the sprint summaries / status notes I paste below. Generate a draft outline I can paste into PowerPoint. One slide per train. ## Rules - If I haven't given you a metric, put "[METRIC TBD]" — don't invent. - Roadblocks should be Bijayta-actionable. "Need PM alignment" is too vague. Better: "Need Bijayta to flag the agentic commerce roadmap gap to Rachel." - Keep each bullet to one line. - End with a 1-slide "decisions I'm asking for" summary across all trains. ## Now ask me for the sprint summaries
What to do this week
| When | Action | Why |
|---|---|---|
| Day 1 | Confirm Copilot seat is live + extensions installed. Sign in. Open Chat. Send one prompt. | Smoke test. Catch any "still pending IT" blockers before you sink time into customization. |
| Day 1 | Run /init in a fresh workspace. Save the generated copilot-instructions.md, then replace it with the SyF-flavored template from this page. | Your instructions file IS the single biggest quality lever. Get this right before adding anything else. |
| Day 2 | Drop the Bijayta prompt file into .github/prompts/bijayta-monthly.prompt.md. Try it on last month's data. | Pick the output with the highest current pain (4hr → target 1hr). Win once, then expand. |
| Day 3 | Ask SyF IT three questions: (1) is data residency enabled on our Copilot org? (2) is Atlassian MCP server allowed? (3) is M365 Copilot Chat live for me? | The first two unlock Tier-3/Tier-4 workflows. The third may already cover 60-70% of your doc-drafting use case today, at no extra cost. |
| Day 4 | Create a personal Copilot Space at github.com/copilot/spaces — name it "Synchrony Pay Vision." Upload your latest deck draft + the routing-logic wiki page. | Spaces replaced Knowledge Bases in Nov 2025. This is the new way to ground Copilot in your project context. Tests the feature on the highest-stakes output you have. |
| Day 5 | Pick ONE PO and draft their Q-coaching log using a prompt file. Save the prompt for the other 3 POs. | PO coaching is 4× quarterly. One prompt that works = 1 hour saved per quarter per PO. |
Open questions to surface with SyF IT
- Is SyF IT going to allow MCP servers?
MCP unlocks the Atlassian integration, which is 80% of the day-job value. If MCP is blocked, you fall back to copy-paste between Confluence and Copilot Chat — workable but slower.
- Is Copilot Spaces visible inside VS Code, or do you have to go to github.com?
Today, Spaces are primarily a github.com surface. VS Code Chat can reference workspace files directly. The story for "use a Space from inside VS Code" is still evolving.
- How do you handle the audit-log gap with SyF security?
GitHub deliberately does NOT retain prompt/response content in audit logs. Some financial compliance teams flag this. Worth surfacing proactively to avoid a surprise block later.
- M365 Copilot Chat vs GitHub Copilot — overlap?
M365 Copilot Chat (Word/Outlook/Teams) covers a lot of the doc-drafting use case with no IDE involved. If both are available, you may use M365 for casual drafts and GitHub Copilot (VS Code) for structured/repeatable workflows where prompt files + Spaces matter.
Sources (research run 2026-05-18)
- VS Code — Using agents in Visual Studio Code
- VS Code — Use custom instructions
- VS Code — Use prompt files
- VS Code — Add and manage MCP servers
- GitHub Docs — Creating Copilot Spaces
- GitHub Changelog — GPT-5.3-Codex base model (2026-05-17)
- GitHub Changelog — Data residency + FedRAMP (2026-04-13)
- GitHub Changelog — Copilot for Jira preview (2026-03-25)
- GitHub Changelog — Knowledge Bases → Spaces migration
- github/awesome-copilot — Community instructions, prompts, agents
- github/spec-kit — Spec-driven development toolkit
- GitHub Blog — Copilot for product + security teams
- GitHub Blog — 5 tips for better custom instructions