Design in Product social media card
← Back to Hub substantive

Cross-Pollination Brief — March 16, 2026

Heavy engineering day on both sides. Piper Morgan closed 9 issues, discovered "extend without verifying" as a shared root cause across 5 simultaneous bugs, and shipped the Action Registry with stub-to-floor routing. Klatch shipped the sidebar redesign (project-first accordion), Rounds 6–8 test suites, formalized 5-layer prompt architecture decisions, and added project settings panel and prompt-debug endpoint.

Key Insights

1. "Extend Without Verifying" — Systemic Root Cause

From: Piper Morgan (Five Whys analysis, commit 454a8445) Relevant to: Klatch

PM's Lead Developer launched a Five Whys analysis on three simultaneous test failures and found a shared root cause: classification was being extended independently of handling. New classifier categories were added, but their handlers were stubs or missing. The system silently absorbed the gap — no errors, just uncovered behavior. Three or more stubbed actions, eight or more pattern overlaps, multiple legacy fallback paths.

Fix: Action Registry (34 classification → handler pairs), stub-to-floor routing (stubs now route to LLM instead of failing silently), multi-intent subsumption filter.

For Klatch: this structural failure mode is available to any routed system. Klatch's five-layer prompt architecture adds conceptual "routes" — if any layer is assembled but never used by the system, it creates a silent gap. The prompt-debug endpoint (added today) helps detect this.

Suggested action (Klatch): Treat the prompt-debug endpoint as a diagnostic tool, not just a developer convenience. When behavior is unexpected, inspect the assembled layers first. Silent gaps in prompt assembly are the Klatch equivalent of PM's stub-to-floor problem.

2. 5-Layer Prompt Architecture — Formalized in Klatch

From: Klatch (docs/plans/, commit e0a7abf) Relevant to: Piper Morgan

Klatch formalized its 5-layer prompt assembly decisions on March 16: Layer 3 (project memory) lives at project level, not channel level. claude.ai imports inject project knowledge into the kit briefing. Channel addendum hidden for standard chats. The prompt-debug endpoint (/channels/:id/prompt-debug) returns all 5 layers with status — allowing inspection of exactly what the model is working with.

PM operates the same 5-layer model conceptually. A debug endpoint that shows the assembled context would help PM agents verify what they're actually receiving in a session. This is the diagnostic capability PM lacks when briefing quality is uncertain.

Suggested action (Piper Morgan): Consider a debug path for PM's context assembly. What would it mean to inspect the assembled prompt for a PM agent session? The equivalent of Klatch's prompt-debug endpoint is a session briefing verification tool.

3. Floor Voice Guidance — "Never Say I Can't"

From: Piper Morgan (CXO memo) Relevant to: Klatch

CXO's guidance for PM's conversational floor: "Never say 'I can't.'" The floor should always engage — for out-of-scope requests, redirect rather than refuse. Friday's contextual fallback copy became test expectations. The framing: a concierge doesn't say "I don't do that." They say "that's not my area, but here's who can help."

For Klatch: when a channel's configured purpose doesn't cover an incoming query, the response pattern matters. The entity prompt and channel addendum should be rich enough to allow Claude to redirect gracefully rather than surface a generic fallback.

Emerging Patterns

Both teams doing systemic cleanup on the same day. PM's Action Registry and Klatch's prompt architecture formalization both address the same underlying risk: routing complexity that grows faster than verification. Both teams chose the same week to address structural debt. Not coordinated — but both projects were at the same point in their development arc.

Background Changes (Noted, Low Priority)

  • Klatch: sidebar redesign (project-first accordion with chat/klatch type column)
  • Klatch: Rounds 6-8 test suites (sidebar redesign, project memory + prompt assembly)
  • Klatch: prompt-debug endpoint (/channels/:id/prompt-debug)
  • PM: #914 GitHub integration tests enabled (keychain + gh CLI), #917 credential leak fixed
  • PM: editorial calendar unified (304 rows, 16 columns, 19 issues fixed)
  • PM: Docs created publish-to-blog skill bridging PM → PM website

Sources Read

Klatch:

  • Git log March 16 — sidebar redesign, Rounds 6-8, prompt architecture, settings panel
  • docs/logs/2026-03-16-0803-daedalus-opus-log.md — sidebar + prompt architecture session

Piper Morgan:

  • docs/omnibus-logs/2026-03-16-omnibus-log.md — Five Whys, Action Registry, floor synthesis, editorial calendar
← Back to Hub