Workbench

Workbench Skill

The Workbench is the project-level semantic memory store for CopilotTraining authoring decisions, patterns, and cross-reference knowledge. It is separate from Copilot Memory (which is personal/cross-project). The Workbench = this repo's institutional knowledge.

Cardinal rule: the Workbench suggests; repo files confirm.

The store is organized as benches (top-level domain folders under memories/) containing drawers (typed .md files inside each bench). Filenames drop the prefixes — a fact drawer in the infra bench is memories/infra/facts.md.

When to Use This Skill

READ — Query the Workbench BEFORE generating content

Invoke after pre-flight gates complete, before starting content generation:

Never query the Workbench during pre-flight. Pre-flight gates (README exists → not archived → TEMPLATE.md → SECTIONS.md) must complete first using live repo files. The Workbench comes after.

WRITE — Update the Workbench at session END

After completing content work, record drawer entries for anything future sessions should know:

Never write during a session mid-stream. Drawer writes happen at session END only — after the work is saved and verified.

Write mechanism: always use replace_string_in_file or multi_replace_string_in_file to edit the drawer .md files directly (e.g. memories/infra/facts.md). Never use the memory tool with /memories/ paths — that routes to Copilot's personal memory service, not the workspace files, and the entries will be invisible in the repo.

What NEVER to Write

• Archived talk content (rationale only, as events)
• Slide deck CSS or Slidev layout details (too volatile, too noisy)
• Secrets, tokens, credentials
• Speculation — only write confirmed facts and decisions
• Anything that duplicates what's in the repo file itself ("copilot-memory is about memory" is useless; "copilot-memory uses a 3-layer model, not 4" is useful)

Bench Reference

Queries and writes should target the appropriate bench:

Drawer Reference

Content Change → Workbench Update Protocol

When you update a content artifact, ask:

1. Does this change a confirmed fact? → Update or add facts entry
2. Does this retire or archive something? → Write events with date and rationale
3. Does this affect a cross-reference? → Update the cross-reference table in slides/facts
4. Did you discover something new that would save time next session? → Write discoveries or advice
5. Topic-specific gate — did this session change anything about this deck's content that an agent reading only the deck would miss? (framing decisions, audience misreads, ordering constraints, why a section was rejected) → Write to the matching topic bench (e.g. agent_architecture/discoveries.md). If the answer is just "we used component X" or "we fixed a Slidev gotcha," that belongs in slides, not the topic bench.

If none of the above apply, no write is needed. Empty topic benches are acceptable — do not pre-create folders or invent thin entries to fill them.

Cross-Reference Table

Maintained in slides/facts. Topics covered in multiple artifacts — check for drift when updating either.

For cross-cutting topics (MCP, agent mode, CLI, custom instructions, memory), use the topics/ bench instead — each has a dedicated drawer with full coverage maps and drift risk notes. The infra/facts table below covers artifact-pair relationships not captured by a topic drawer:

• MCP transport: 05-mcp-servers, mcp-apps → see topics/mcp.md
• Instructions files: 01-instructions, copilot-primitives → see topics/custom-instructions.md
• Agent sessions: vscode-latest, 06-custom-agents → see topics/agent-mode.md
• Memory layers: copilot-memory, 05-mcp-servers → see topics/copilot-memory.md

Example Workflow (Slide Generator)

1. Pre-flight completes:
   README exists ✓ → not archived ✓ → TEMPLATE.md ✓ → SECTIONS.md ✓ → deck.recipe.yml ✓

2. QUERY the Workbench (now):
   "What patterns or gotchas exist for [talk-slug]?"
   → Returns: "cockpit wrapper eliminates HTML balance errors" (discoveries)
   → Returns: "Thank-you slide pattern: blur-2xl double-logo, gradient h1" (advice)
   Apply these to generation.

3. Generate slides.

4. Session END — WRITE drawer entry:
   "Discovered: agentic-sdlc opener stat (10-15 features/day) works best as hero number on slide 2"
   → agentic_sdlc / discoveries

The Workbench Is NOT

• ❌ A replacement for reading repo files — always confirm against source
• ❌ A wake-up injection context (don't front-load sessions with workbench dumps)
• ❌ Copilot Memory (that's personal/cross-project preferences)
• ❌ A content archive (use git history for that)