Failure Modes — Quick Reference

When a chain breaks, walk this list before rewriting anything. Three failure modes account for nearly every break in nearly every workflow.

How to Use This Sheet

Identify the leg that broke (sender → payload → receiver).
Walk failure modes 1 → 2 → 3 in order.
Apply the matching repair.
Update your three-spec-worksheet.md for that workflow.
Re-run the leg.

If you try this and the chain still breaks, the issue is upstream — go back one leg.

Failure Mode 1 — Unspecced Handoff

What it looks like: You wrote a prompt or sent a message but never named the leg. The chain runs but the output drifts every time. Different result on Monday vs. Friday. Sometimes good, sometimes generic, no pattern.

Why it happens: You optimized for the prompt or message in isolation. The leg around it has no spec, so the model (or human) on the receiving end orients differently every time.

Repair — 3 steps:

Identify which kind of leg it actually is — human↔human, human↔AI, or AI↔AI.
Run the matching spec on it:
- Human leg → Communication Diamond (5 facets)
- Human↔AI leg → AI Spec (4 slots)
- AI↔AI leg → Handoff Contract (3 slots)
Save the spec next to the workflow. Re-run.

Frequency: Roughly 80% of chain breaks fix at this step. Try this first.

Failure Mode 2 — Wrong Spec on the Wrong Leg

What it looks like: You DID write a spec, but it's the wrong one for the leg. Two common shapes:

You applied the AI Spec (with anti-patterns and audience-for) to a human leg → output reads cold and robotic to your client.
You applied the Communication Diamond (with motivation blockers and ownership) to an AI leg → the model ignores half of it because it doesn't process motivation the way humans do.

Why it happens: The specs aren't interchangeable. A leg's sender and receiver determine which spec applies. Most agents pick whichever spec they remember most, regardless of leg type.

Repair — 3 steps:

Name the sender and receiver of the broken leg explicitly.
- Both human → Diamond
- One human, one AI → AI Spec
- Both AI (no human in the middle) → Handoff Contract
Pull the spec for that leg and rewrite it from scratch using the right facets/slots.
Run the leg. The output should immediately feel "in-pocket" again.

Tell: if rewriting the prompt didn't help but rewriting the spec with the right shape did, this was your failure mode.

Failure Mode 3 — Frozen Spec

What it looks like: The spec was right six months ago. It's still in your folder. The output is now off — drifting tone, missing context, ignoring constraints that used to land. Nobody touched the spec, but the world around it moved.

Why it happens: Your business changes. Markets shift. Models upgrade. Voices drift. Assistants change roles. The spec was a snapshot; it doesn't auto-update.

Repair — 1 step (and one habit):

Right now: open the spec and audit it against current reality. Update tone, vocabulary, audience, anti-patterns, and any examples baked in.
Habit: every spec gets a 30-day re-orient date in your calendar. Not a sticky note. A real calendar event with a 15-minute review.

Tell: if the spec hasn't been touched in 30+ days and the output is drifting, this is the failure mode. Don't rewrite from scratch — re-orient the existing spec.

Diagnosis Tree (for the back of your monitor)

┌─────────────────────────────────────────────────────────┐
│ A workflow gave you bad output.                         │
│                                                         │
│ Q1: Which LEG broke?                                    │
│   → name sender, payload, receiver for that leg.        │
│                                                         │
│ Q2: Was there a SPEC on that leg?                       │
│   → No  → Failure Mode 1 (Unspecced). Run matching      │
│           spec. 80% of breaks end here.                 │
│   → Yes → continue.                                     │
│                                                         │
│ Q3: Was it the RIGHT KIND of spec for that leg?         │
│   → No  → Failure Mode 2 (Wrong spec). Identify         │
│           sender + receiver, apply matching spec.       │
│   → Yes → continue.                                     │
│                                                         │
│ Q4: When did you last RE-ORIENT the spec?               │
│   → 30+ days ago → Failure Mode 3 (Frozen). Audit       │
│                    against current reality. Add         │
│                    calendar event.                      │
│   → Recent      → the issue is upstream. Go back        │
│                    one leg in the chain. Repeat Q1.     │
└─────────────────────────────────────────────────────────┘

When None of These Apply

If you've walked all three failure modes and the chain still breaks, the issue is one of:

Upstream leg. The leg you're staring at is fine; the leg before it is feeding it bad payload. Move one leg upstream and run the diagnosis again.
Missing leg. You think the chain has 3 legs but it actually has 4 — one is implicit and unspecced. Map the chain end-to-end. Look for handoffs you didn't notice.
Wrong tool entirely. Rare, but real. The leg has the right spec and runs correctly; the tool just isn't the right tool for the job. This is the last hypothesis to test, not the first.

Class: Innovation Lab · The Communication Loop That Doesn't Break Companion files: three-spec-worksheet.md, chain-anatomy-cheatsheet.md