Draft-First Support Agents Keep Humans in the Loop
draft-first support flow maps to joelclaw's agent-loop reviewer gate and Skill Recordings Front triage
This Loom is a rough walkthrough of a new Skill Recordings support agent repo: Bun, Biome, Turborepo, apps for Front, Slack, and web, plus Inngest workflows behind incoming support email.
The useful shape is draft first, approval second, autonomy later. The agent reads Front conversations, skips things it shouldn’t answer, drafts replies it clearly marks as AI-generated, and routes approval signals through Slack. If a human sends the draft as-is, the system can treat that as confidence. If the human edits or deletes it, that’s training data instead of a silent failure.
The clever bit is that the support agent isn’t just chat glued to an inbox. It has a CLI, Stripe Connect read access through connected accounts, product adapters installed in AI Hero and Total TypeScript, a typed Front API layer, content search, and a semantic memory system for decisions and answered questions.
The build process is also the artifact: a split PRD, phase files, local docs, and coding-agent rules created by mining Claude Code and OpenCode transcripts for corrections. That’s the part worth stealing for joelclaw: every agent mistake should become a rule, skill, or test so the system gets less annoying over time.
Key Ideas
- A support agent can start safely by creating Front drafts instead of sending replies directly.
- Slack approvals and reactions become lightweight feedback signals for future confidence scoring.
- Stripe Connect lets one top-level account query connected product accounts without copying every product’s API key.
- Product-specific SDK methods like user lookup, purchase lookup, purchase transfer, and magic-link generation create a clean adapter boundary for support automation.
- A CLI gives both humans and agents a shared debugging surface for Front, Inngest, and support workflows.
- A split PRD plus phase files keeps Claude Code and OpenCode from drowning in one giant requirements blob.
- Mining coding-agent transcripts for corrections turns mistakes into durable rules, skills, and tests instead of repeating the same bullshit.
- The human stays in the loop until trust is earned, which matches the reviewer-gate shape of agent loops.