💡 Natural Language Custom Workflow Builder: Voice-Defined Reusable Automation for Non-Technical Users

[github-actions[bot]]

Summary

Users describe reusable multi-step workflows in natural conversation with the avatar, which guides them through defining triggers, steps, and outputs. The resulting workflow is stored as a named template that can be invoked by voice command in future sessions, empowering non-technical users to build their own automation without code or visual builders.

Market Signal

Natural language interfaces are replacing visual workflow builders in 2026. Non-technical users face a gatekeeping bottleneck where IT controls every process change — Kissflow, Cflow, and AIMultiple research all identify "natural language workflow definition" as the breakthrough trend for democratizing automation. ChatGPT Work (launched July 2026) can execute tasks but doesn't support user-defined reusable workflow templates — this is a differentiation opportunity.

User Signal

Existing ideas #244 (Workflow Template Gallery) provides curated starting points and #233 (Natural Language Automation Rules) covers scheduling, but neither addresses user-created custom workflows. Non-technical users (Journey 1: Sarah the PM) need to define their own recurring processes beyond pre-built templates — for example, "every Monday, summarize what changed in the project this week and draft a status update."

Technical Opportunity

Claude Agent SDK's subagent capability can decompose a natural language workflow description into structured steps. The agent generates a reusable template (JSON schema with trigger, steps, output format, required MCP tools) stored in the workspace. Future invocations parse the template and execute via the same SDK pipeline. The AgentBackend abstraction layer supports this without architectural changes. The plan preview component (FR44) provides the confirmation UX.

Assessment

Dimension	Score	Rationale
Feasibility	med	Requires workflow template schema design and a robust natural language → structured steps pipeline. The confirm-plan pattern (FR44) already provides the interaction model.
Impact	high	Transforms TalkTerm from a tool runner into a personal automation platform for non-technical users — a significant differentiation from both ChatGPT Work and CLI tools.
Urgency	med	ChatGPT Work's July 2026 launch increases competitive pressure, but this is a Phase 2+ feature that builds on the MVP workflow execution foundation (Epic 6).

Adversarial Review

Strongest objection: Users might create ambiguous or overly complex workflow descriptions that fail unpredictably. Quality control for user-defined workflows is inherently harder than curated templates.

Rebuttal: The avatar-guided refinement loop mitigates ambiguity: when the user describes a workflow, the avatar plays it back step-by-step, asks clarifying questions, and shows a plan preview (FR44) before saving. A "test run" option lets users validate before committing. This is the same confirm-plan pattern already built into the architecture. Complexity limits can be enforced (max steps, required MCP tool availability checks).

Suggested Next Step

Design the workflow template schema (trigger, steps, output format, required MCP tools) and prototype a single-workflow creation flow using the existing plan preview component. Start with a constrained scope: templates that chain 2-5 agent actions with text output.