skills/design-optimal-ai-workflow/references/provider-candidates.md

# Provider candidates, not automatic assignments

Official-source check: 2026-09-14. These are conditional responsibility candidates, not performance rankings or proof of local access. Recheck changing features in the actual target surface. Use the shared [capability states](../../../shared/expert-system/provider-capability-routing.md); no external provider was connected, configured or runtime-tested by writing this profile.

| Candidate | Consider for | Boundary / verify before use |
|---|---|---|
| ChatGPT Chat | Research synthesis, reasoning, planning, product/business discussion, supplied multimodal material | Actual tool and modality access; no assumption of local filesystem access or installed skill invocation |
| ChatGPT Work | Multi-step artifact work and permitted local/connected-resource workflows | Local versus cloud environment plus file, shell, computer-use, browser and connector capabilities must each be checked; one does not prove another |
| Claude Cowork | Long-running, multi-step knowledge work across approved files, folders, browser or connected apps; document and execution-package compilation or evidence review | Verify actual product access, selected workspace/mount, read/write mode, connectors and isolated versus local-host execution; Cowork output is derived/proposed unless explicit authority makes it canonical |
| Codex / Claude Code | Repository changes, runtime investigation, Git and tests | Actual repository scope, installed runtime, commands, permission mode and evidence; neither gets product authority by default |
| Claude Design | Visual exploration and interactive prototypes | Exploration output is proposal/proof, not automatically approved Product Authority or production code |
| Blender | Authored 3D source work when geometry/material control is needed | Check actual installation, asset/format compatibility, exporter and runtime proof; do not force 3D where simpler representation meets the target |
| Specialized plugin/connector or deterministic tool | A specific supported action with better fit and fewer handoffs | Inspect actual operation, authorized destination, configuration and representative test, not catalog presence alone |

Evidence for documented scope: [ChatGPT / Work / Codex overview](https://learn.chatgpt.com/), [Work local/cloud overview](https://learn.chatgpt.com/docs/get-started-with-work), [Work local security and capability boundaries](https://learn.chatgpt.com/docs/enterprise/chatgpt-work-local-security), [Claude Cowork environment and file boundaries](https://www.anthropic.com/engineering/how-we-contain-claude), [Claude Cowork workflow overview](https://www.anthropic.com/webinars/cowork-workshop-foundations), [Claude Code operation](https://code.claude.com/docs/en/how-claude-code-works), [Claude Design](https://claude.com/product/design). These sources establish product descriptions, not comparative superiority, current feature rollout or the user's account entitlements. Blender remains a DESIGNED candidate from the owner's requested workflow: retrieval of the official manual and modeling feature page failed on this check. Current Blender capability/version/export support is not verified; check it before selection.

For concrete Claude Code mechanisms use the existing [Claude Code provider profile](../../orchestrate-coding-agent-execution/references/providers/claude-code.md); do not copy its commands into shared policy. Model names/prices, plan limits, automatic updates and cross-account skill availability are not frozen into this profile. A manual artifact handoff is the fallback when a required connection is unverified; it still needs permitted data transfer and recipient access.