Build note
AI Workflow Coach: repeated AI coding failures, backed by local evidence
Heavy AI coding users do not need another transcript browser. They need evidence that explains why sessions keep failing in the same way.
The problem
AI coding sessions fail in patterns. The agent claims completion after a failed tool call. The user repeats instructions that should live in the repo. The session drifts because nobody stops to check what happened before closeout.
Why transcript search is not enough
Search can find a bad turn. It does not explain whether the failure repeats, whether the cause is user prompt shape, agent behavior, repo governance, or tool setup, and where the smallest durable fix belongs.
The product shape
AI Workflow Coach is a CLI plus an agent skill. The CLI parses local sessions and emits candidate patterns. The skill reads the evidence, cites a real quote, attributes carefully, and gives one change to try next session.
What v1 avoids
No dashboard, no workflow framework, no marketplace, no automatic edits, and no persistent ledger. The first useful version is local and read-only. Its job is to make the next agent session cleaner, not to become the operating system for all work.
What I want the output to change
The note should point to a smaller instruction file, a tighter skill, a targeted hook, a better plan document, or a changed prompting habit. If it cannot name the evidence and the surface, it should abstain.
The design constraint
The CLI owns facts: session discovery, parsing, candidate recurrence, and validation. The agent owns judgment: attribution, teaching, and the smallest useful intervention. That split is what keeps the product from turning into another pile of AI advice.