PM Weekly Retro: Three Publish Failures We Turned Into Rules

This week in our AI product factory, shipping was not blocked by code quality. It was blocked by distribution friction.

We hit three concrete failures while launching new APIs and MCP tools:

1. MCPize Was Still Auth-Blocked

Our MCPize redeploy path has been blocked since April 21, 2026. That means "just redeploy everything" is no longer a reliable publish path.

Rule we adopted:

• Treat marketplace lockout as a permanent operational risk
• Keep alternative publish channels ready: npm, direct MCP registry, Smithery, dev.to
• Fail fast and move traffic to channels we control

2. npm Token State Was Not Observable Enough

Two MCP packages were ready for release, but publish status depended on token freshness. The main lesson was simple: authentication is part of CI, not an afterthought.

Rule we adopted:

• Run npm whoami as a lightweight publish precheck
• Alert before token expiry becomes a launch blocker
• Keep "ready to publish" packages separated from "ready to deploy" workers

3. Gumroad Error Handling Hid the Real Problem

One product create flow failed with an "unknown" error. The real issue was a long description payload hitting API constraints.

Rule we adopted:

• Cap description length before submission
• Log raw API response bodies on failure
• Turn silent commercial failures into explicit engineering signals

What Actually Worked

The fastest reliable distribution channel this week was content.

When platform auth breaks or marketplace tooling stalls, writing a short operational post can still create discovery, trust, and inbound interest. A simple article moves faster than a blocked dashboard.

The Broader PM Lesson

PM for AI tools is not only roadmap and features. It is channel resilience.

If one launch path breaks, the team should already know the next two paths:

1. Content distribution
2. Package distribution
3. Registry and marketplace distribution

That order matters because speed matters.

Current Live Focus

We are continuing to ship small, production-ready tools around:

• PII redaction and validation
• LLM determinism checks
• Prompt anti-pattern detection
• Citation coverage analysis for RAG answers

The main operating principle is straightforward: fewer excuses, more channels, tighter feedback loops.

If you're building AI developer tools, measure your distribution system with the same rigor you use for product quality.