💡 Outcome-Driven Workflows: Avatar-Guided Success Criteria with Visual Scorecard

[github-actions[bot]]

Summary

Integrate Claude's new Outcomes feature to let users define what "good" looks like before starting a workflow, and display a visual quality scorecard after completion showing how well the output met their goals. The avatar asks "What would make this brainstorming session a success for you?" and uses an independent grader to evaluate output quality, retrying if the result falls short.

Market Signal

Anthropic shipped the Outcomes feature in Claude Managed Agents (May 7, 2026). It uses independent grading — a separate evaluator scores agent output against user-defined criteria, preventing the agent from marking its own homework. Gartner's 40% enterprise AI agent adoption prediction (2026) requires quality assurance mechanisms that non-technical users can trust.

User Signal

TalkTerm's target users (non-technical knowledge workers) need confidence that AI output is good enough to use. Trust is the #1 barrier to adoption. FR20-22 handle confirmation before actions, but nothing currently validates output quality after completion. The audit trail (FR27-29) logs what happened but not whether it met the user's goals.

Technical Opportunity

The Outcomes rubric can be stored alongside session metadata in SQLite (main/storage/sessions.ts). The grader evaluation runs as a post-workflow step in claude-sdk-backend.ts. The visual scorecard is a new display mode in renderer/components/display/ — a natural extension of the existing multi-mode panel (FR45). The confirm-plan pattern (FR20) already establishes the avatar-guided pre-workflow conversation flow.

Assessment

Dimension	Score	Rationale
Feasibility	med	Depends on Outcomes API availability for Agent SDK users. Core UX patterns (pre-workflow conversation, display modes) already exist in the architecture.
Impact	high	Closes the trust gap for non-technical users. Independent grading provides quality assurance without requiring users to evaluate AI output themselves.
Urgency	high	Outcomes just shipped (May 2026). Quality assurance is critical for enterprise adoption (Gartner 40% prediction).

Adversarial Review

Strongest objection: Adding a "what does success look like?" conversation before every workflow adds friction. Non-technical users may find it annoying or not know how to define criteria.

Rebuttal: The conversation is brief and natural — "What are you hoping to get out of this?" is a question any good consultant asks. If the user says "just get started," the avatar skips the rubric and uses sensible defaults. The scorecard at the end is optional but builds trust over time. The grader cost is marginal — one API call to evaluate output vs. criteria.

Suggested Next Step

Evaluate Outcomes API availability for Agent SDK users. Design the pre-workflow rubric conversation flow (avatar prompts, skip option). Prototype the visual scorecard display mode component.