This folder defines how Copilot and agents behave in this repository.
This layout follows current GitHub and VS Code guidance:
- Repo-wide instructions:
.github/copilot-instructions.md - Path-specific instructions:
.github/instructions/*.instructions.mdwithapplyTofrontmatter - Custom agents:
.github/agents/*.agent.md - Agent memory/rules support:
AGENTS.mdat repo root
Important:
- VS Code can combine multiple instruction files for chat context.
- Instruction order is not guaranteed when multiple files apply.
- Keep files single-purpose and non-overlapping to avoid conflicts.
copilot-instructions.md: global behavior, tool routing, execution contractPROMPT_SECURITY_TEMPLATES.md: reusable security/privacy/prompt-injection blocksagents/agent-live-web.agent.md: specialized live-web custom agentinstructions/*.instructions.md: path-scoped implementation rulesinstructions/README.md: authoring and maintenance guide for scoped rulesskills/web-works/SKILL.md: website workflow policyskills/web-works/web-task.schema.json: strict schema for JSON-driven web tasksskills/web-works/web-task.template.json: reusable template for website tasksskills/web-works/PROMPTS.md: prompt pack (master/resume/deep/turbo)
- Global behavior:
copilot-instructions.md - Security snippets:
PROMPT_SECURITY_TEMPLATES.md - Runtime constraints:
instructions/*.instructions.md - Task workflow:
skills/web-works/SKILL.md - JSON task contract:
skills/web-works/web-task.schema.json
.github/**andAGENTS.mdare protected by default during normal tasks.- Agents should not modify governance files unless the user explicitly requests governance updates.
- Need global behavior change: edit
copilot-instructions.md - Need path-specific rule: add or edit
instructions/*.instructions.md - Need security prompt hardening: edit
PROMPT_SECURITY_TEMPLATES.md - Need different agent persona/tool list: edit
agents/*.agent.md - Need website workflow change: edit
skills/web-works/SKILL.md - Need JSON schema change: edit
skills/web-works/web-task.schema.json
- Copy
skills/web-works/web-task.template.json. - Fill
start_url,objective,success_criteria, andsteps. - Use
site_profile: "generic"for most websites andsite_profile: "whatsapp-web"for WhatsApp Web. - Keep
auto_send_allowed: falseunless explicitly approved. - Execute one action per step with strict verification.
- Use VS Code ownership mode (
vscode) for browser work. - Validate task JSON against
web-task.schema.json. - Run one tool action at a time and verify result.
- Ask confirmation before irreversible side effects.
- Report blockers with exact tool/selector/evidence.
- GitHub custom instructions:
- VS Code instructions model:
- VS Code custom agents:
- OpenAI safety guidance: