Tune how the model interacts with the user in VS Code Copilot — verbosity, autonomy, trust level, output format, feedback capture, status reporting, error communication, and per-agent personality. Trigger on: "too verbose", "too quiet", "stop asking", "ask more", "trust calibration", "output style", "personality", or any complaint about how the model communicates. Also trigger when someone says the model is "annoying", "slow", "over-explains", or "doesn't explain enough".
Tune the interaction between user and model — the "feel" of working together. Good ergonomics mean the model communicates at the right level, makes the right decisions autonomously, and captures feedback to improve over time.
A model that over-explains wastes time. One that under-explains makes mistakes silently. One that asks permission for everything kills flow. One that never asks causes disasters. Ergonomics calibrate the sweet spot for each user and project.
Different trust levels for different domains:
grep -i 'verbose\|concise\|terse\|explain\|confirm\|autonomous\|style\|tone' .github/copilot-instructions.md 2>/dev/null
for f in .github/agents/*.agent.md; do
echo "=== $(basename $f) ==="
# Look for personality/style instructions in body
grep -i 'verbose\|concise\|terse\|explain\|style\|tone\|brief' "$f" 2>/dev/null
done
ls .github/memory/feedback_*.md 2>/dev/null
"How would you describe the ideal interaction style? Pick one or describe your own:"
Add ergonomics guidance to copilot-instructions.md:
## Interaction Style
- Execute tasks autonomously. Report results, not plans.
- Terse output. One-line confirmations for routine work.
- Only ask for confirmation on destructive or irreversible operations.
- Make decisions. Don't enumerate options unless the choice is genuinely ambiguous.
- No filler: skip "Let me...", "I'll now...", "First, I'll...".
- Errors: state what happened, why, and what to do. No apologies or speculation.
## Interaction Style
- Auto-execute routine tasks (format, lint, simple edits) without asking.
- Confirm before: git push, deploy, delete files, modify shared config.
- Concise output: state what was done, not what was considered.
- Surface non-obvious decisions with a recommendation and one-line rationale.
- Errors: what happened, why, how to fix. Include the relevant error output.
## Interaction Style
- Explain your approach before executing multi-step tasks.
- Confirm before modifying files. Show what will change.
- Include brief explanations of why patterns are used.
- Add code comments for non-obvious logic.
- Summarize what was done at the end of each task.
Different agents benefit from different styles:
You are direct and opinionated. Point out issues concisely.
Don't hedge — if something is wrong, say so.
Format: issue → location → fix suggestion (one line each).
Present analysis in structured format: context, options, recommendation.
Include trade-offs for each option.
Use tables for comparisons. Keep prose under 4 sentences per point.
Execute without narrating your plan. Report what you did in 1-2 lines.
Only ask questions when genuinely blocked (missing info, ambiguous requirement).
Feedback memories record corrections and confirmations so the model improves over time.
When the user says "no not that", "don't", "stop doing X":
Create a feedback memory:
---