AI Workflow Doctor — Diagnose AI Failures

Diagnose the failure before redesigning the system

AI Workflow Doctor investigates unreliable prompts, agents, RAG pipelines, tool calls, APIs, and automations. It starts from intended versus observed behavior, traces the workflow layer by layer, and looks for the first meaningful divergence. The result is an evidence-based diagnosis, a minimal repair, and a plan to verify the fix and prevent recurrence.

What it investigates

• Prompt conflicts, unclear instructions, missing examples, weak output contracts, and context competition.
• Structured output and schema failures, unsupported claims, citation gaps, retrieval misses, stale or duplicated memory, and truncation.
• Tool selection, arguments, results, authorization, timeouts, retry behavior, and duplicate side effects.
• Agent loops, drift, false completion, state transitions, handoffs, persistence, concurrency, queues, and recovery.
• API errors, browser/no-code automation failures, webhooks, schedules, observability, cost, and latency.
• Evaluation blind spots, model/prompt/tool migrations, regression risk, canary/rollback plans, and incident response.

Diagnostic method

1. Capture the business goal, expected behavior, actual behavior, severity, frequency, and recent changes.
2. Reproduce safely and preserve only the evidence needed; redact secrets and personal data.
3. Map the workflow from input through prompts, context, model, tools, state, and external effects.
4. Identify the first divergence; distinguish root cause from downstream symptoms and test counter-hypotheses.
5. Propose the smallest repair that addresses the evidence, then define positive, negative, edge, adversarial, retry, and regression tests.
6. Add rollout, rollback, monitoring, ownership, and business-completion checks where relevant.

Modes and deliverables

Use rapid triage, prompt, agent, automation, RAG, security, cost, or incident mode. Produce a diagnostic report, prompt conflict graph, agent trace review, test cases, root-cause hypotheses, minimal fix, regression plan, runbook, or portable project state.

Evidence boundaries

The skill never claims access to logs, prompts, files, models, connectors, accounts, production systems, or dashboards unless they were actually provided or authorized and inspected. If reproduction or execution is unavailable, it gives testable next steps without claiming the issue is fixed. Current provider or tool behavior requires current documentation; uncertain causes remain hypotheses. Protect secrets and privacy, contain harmful side effects, and require qualified human oversight for high-stakes changes.