Skip to main content
The Hard Parts.dev
FM · Failure Modes Issue 01

Failure Modes

Named failure patterns that recur in software engineering: how they start, how they escalate, what they look like at each stage, and what good responses look like.

Entries

31

Classes

06

Classes

  • planning
  • people
  • technical
  • process
  • leadership
  • ai

The catalog is grouped by where the pattern lives: planning, people, technical, process, leadership, AI-specific. Each entry carries a severity rating shown two ways: as a five-step bar on the card and as the card’s own weight. The fill darkens through five grays, from paper at low to near-black at critical. Color is reserved for category identity; severity reads purely as gravity. Scan the grid for weight first, then read the names.

  • Severity key

    • low
    • medium
    • medium-high
    • high
    • critical

    Chip = card fill on the grid.

  • Frequency

    How often this pattern actually shows up in practice: from rare one-offs to near-universal.

    increasing Not a point on the scale. A trend. Flags patterns whose prevalence is rising (often AI-era).

    • rare
    • occasional
    • common
    • very common
    • universal
  • Recovery

    How hard it is to climb back out once you are in it: tactical fix vs. structural teardown.

    • easy
    • medium
    • medium-hard
    • hard
    • very hard
  • Confidence

    How sure we are the pattern is real and consistent: provisional vs. repeatedly observed.

    • low
    • medium
    • medium-high
    • high
Class 01

Planning

02 entries
Class 02

People

06 entries
Class 03

Technical

06 entries
Class 04

Process

02 entries
Class 05

Leadership

06 entries
Class 06

Ai

09 entries
FM-013 ai

Synthetic Velocity

Output volume rises sharply while true understanding, maintainability, and durable progress do not.

Freq · increasing
FM-014 ai

Silent Model Drift

Model behavior changes materially in production before the organization notices or responds effectively.

Freq · increasing
FM-015 ai

Autocomplete Architecture

Teams accept AI-suggested structures faster than they understand or own them, embedding design decisions nobody made consciously.

Freq · increasing
FM-016 ai

The Benchmark Mirage

Model selection or evaluation is guided by benchmark performance that does not reflect real production behavior.

Freq · common
FM-017 ai

RAG Without Ground Truth

A retrieval-augmented system is built and deployed before source quality, citation reliability, and answer validation are established.

Freq · increasing
FM-018 ai

Prompt Ops Chaos

Prompts, model settings, and hidden instructions change without version control, making system behavior unpredictable and undebuggable.

Freq · common
FM-027 ai

Eval Goodhart

Internal evaluation sets become optimization targets rather than honest capability measures, producing models or prompts that score well but behave poorly in production.

Freq · increasing
FM-028 ai

Context Window Hoarding

Teams fill context windows maximally with documents, history, and examples without understanding what actually helps, leading to unpredictable behavior, high cost, and debugging nightmares.

Freq · increasing
FM-029 ai

Human-in-the-Loop Decay

Human review steps designed to catch AI errors are gradually skipped as volume increases and confidence grows, removing oversight before the risk does.

Freq · increasing