Auto-Reply Workflow
Instead of triaging replies by hand every day, wire a durable graph8 workflow: reply lands -> LLM skill drafts a contextual follow-up -> draft is ready for review/send. Built conversationally, validated, and delivered DISABLED so the user flips it on deliberately.
Note on naming: this skill uses graph8’s g8_skill_* tools, which author LLM/API skill records that workflow Action nodes execute. Those are a different thing from Skill Library entries like this file (instructions you install into your agent). This skill composes both worlds.
When to use
- “Automatically draft follow-ups when someone replies”
- “Build a workflow that handles campaign replies”
- “Stop making me triage manually every morning”
For one-off manual triage, use triage-and-reply-inbox instead.
Prerequisites
- Activate the two tool families:
g8_tool_search("workflow")andg8_tool_search("skill"), or call viag8_execute.
Steps
- Discover the building blocks.
g8_workflow_list_node_typesto see available triggers and actions;g8_workflow_describe_node_typeon the reply trigger to learn its output fields (sender, subject, body, thread context). Do not guess field names; read them. - Author the LLM skill.
g8_skill_create_llmwith a prompt template that consumes the trigger’s fields as single-brace placeholders, e.g.{sender_name},{reply_body}. Single-brace{var}is the contract the workflow builder detects; double-brace{{var}}is treated as literal text. Use dry_run first, show the user the prompt, then persist. - Design the graph. Trigger (reply received, scoped to the right campaign/mailbox) -> Action node running the LLM skill with input mappings wiring trigger outputs to the skill’s placeholders -> delivery step per the user’s preference (draft-for-review is the safe default; fully-auto send only if the user insists and understands the risk).
- Create disabled.
g8_workflow_createwithenabled=false(dry_run preview first, then persist). A workflow that has never been reviewed must not be live. - Validate.
g8_workflow_validate; fix any missing references or unmapped placeholders it reports. An LLM skill placeholder with no input mapping means the model gets literal{var}text at run time. - Walk the user through it. Show the graph in plain language: “When X happens, Y drafts a reply using Z context, and it lands as a draft for your review.” State the per-execution credit cost.
- Enable on explicit approval only.
g8_workflow_updatewithenabled=trueafter the user confirms. Suggest a test: send a reply to the campaign from a test address and watch the workflow produce a draft.
Example prompts
- “When someone replies to the launch campaign, draft a follow-up in my voice for review.”
- “Automate first-touch responses to demo requests, but let me approve sends.”
- “Build a reply-handling workflow for the RevOps sequence.”
Troubleshooting
- Validation reports missing references: re-check input mappings; every
{placeholder}in the LLM skill needs a mapped source from the trigger. - Skill saved but variables not detected: you used double braces; recreate the prompt with single-brace
{var}. - Tool not visible: both families are gated;
g8_tool_search("workflow")/g8_tool_search("skill")first.
Install this skill
Add the graph8 skill marketplace once, then install auto-reply-workflow from the
/plugin menu:
claude plugin marketplace add graph8-com/skill-libraryOr install manually by saving the skill file into your project:
mkdir -p .claude/skills/auto-reply-workflowcurl -o .claude/skills/auto-reply-workflow/SKILL.md https://raw.githubusercontent.com/graph8-com/skill-library/v0.1.0/skills/auto-reply-workflow/SKILL.md-
Save the skill file into a folder named
auto-reply-workflow:Terminal window mkdir -p auto-reply-workflowcurl -o auto-reply-workflow/SKILL.md https://raw.githubusercontent.com/graph8-com/skill-library/v0.1.0/skills/auto-reply-workflow/SKILL.md - In claude.ai or Claude Desktop, open Settings → Capabilities and choose Upload skill.
-
Select the
auto-reply-workflowfolder. The skill activates automatically when a conversation matches it.
Save the skill as a Cursor project rule:
mkdir -p .cursor/rulescurl -o .cursor/rules/auto-reply-workflow.mdc https://raw.githubusercontent.com/graph8-com/skill-library/v0.1.0/skills/auto-reply-workflow/SKILL.md
Connected over the graph8 MCP instead? No install needed — run
g8_tool_search("library") and the same playbooks surface as callable tools.