AI coding field notes

The strange, funny, and expensive ways AI debugging goes wrong

A living field guide to invented APIs, cargo-cult fixes, version confusion, premature victory laps, and the prompts that help pull a coding assistant back toward evidence.

NEW EVERY DAY

A date-based AI debugging roast now rotates automatically after midnight in Asia/Shanghai, with six previous notes kept on the page.

DAILY AI ROAST

2026-10-02 · #20727

The AI added logging to every branch.

The bug is still hidden, but it now produces a newsletter.

Observability flood

HUMAN TRANSLATION

Instrumentation needs a hypothesis.

One original debugging joke is selected from the maintained ContextClean library using the current date in Asia/Shanghai. It remains stable for the day and changes automatically after midnight.

Previous six field notes

2026-10-01

The null pointer now has an excellent name.

2026-09-30

The missing environment variable remained shy in 40,000 files.

2026-09-29

The product requirements were relieved to learn they had been wrong.

2026-09-28

The function was lonely, so it received a network boundary.

2026-09-27

Now the log, the parser, and the developer all contain unmatched groups.

2026-09-26

It recommended defining it. Computer science advanced another step.

Problem atlas

Eight recurring failure patterns and how to challenge them

PATTERN 01

API fan fiction

What it looks like: The answer uses a plausible method, option, or package that does not exist.
Why it happens: Language models optimize for plausible continuations, not verified library inventories.
Counter-prompt: Ask it to cite the exact package version and documentation surface, or verify the symbol in your installed types.

PATTERN 02

The dependency bonfire

What it looks like: The first recommendation is deleting lockfiles, caches, and node_modules.
Why it happens: Resetting state sometimes works, so it becomes an overused generic escape hatch.
Counter-prompt: Ask which evidence points to stale state and what lower-cost check can confirm that hypothesis.

PATTERN 03

Symptom whack-a-mole

What it looks like: The patch adds guards or optional chaining where an impossible value appears.
Why it happens: The visible exception is easier to patch than the earlier state or data-flow error.
Counter-prompt: Ask where the invalid value was first introduced and require the answer to trace it backward.

PATTERN 04

Framework cargo cult

What it looks like: The answer adds useEffect, dynamic import, or client-only rendering to silence a warning.
Why it happens: Common fixes from unrelated examples are applied without checking the component boundary.
Counter-prompt: Require an explanation of the server value, client value, and first render that differs.

PATTERN 05

Version amnesia

What it looks like: The solution matches an older or newer API than the project actually uses.
Why it happens: Training examples span incompatible versions and documentation generations.
Counter-prompt: Put exact runtime and package versions near the top of the request.

PATTERN 06

Refactor inflation

What it looks like: A one-line bug receives a new abstraction, helper module, and architecture proposal.
Why it happens: Longer answers can appear more complete even when they increase change risk.
Counter-prompt: Request the smallest safe fix first, then ask for optional cleanup separately.

PATTERN 07

Evidence blindness

What it looks like: The answer never mentions the assertion difference, exit code, or first application frame.
Why it happens: Too much surrounding log text dilutes the highest-signal evidence.
Counter-prompt: Lead with a reviewed excerpt and explicitly ask the model to cite evidence for each claim.

PATTERN 08

Premature victory lap

What it looks like: The response says the issue is fixed without a test, command, or observable result.
Why it happens: Generating code and verifying behavior are different tasks.
Counter-prompt: Require one verification command and define the expected output before applying the change.

AI says

What it often means

This should resolve the issue.

I generated a plausible patch but did not run it.

Make sure your dependencies are up to date.

I have not isolated which dependency matters.

There may be a caching issue.

The evidence did not support my first guess.

You are absolutely right.

The previous confident answer was wrong.

For production, consider adding robust error handling.

The root cause is still unresolved, but try/catch is available.

This is a common issue.

I have seen similar words in unrelated examples.

Painfully accurate poll

Which AI coding habit wastes the most time?

This lightweight vote is stored only in your browser. Use the community link below when you want a public, replyable discussion.

Answer smell detector

Does this AI answer deserve to touch your code?

The answer does not quote or reference the actual error.It recommends a major upgrade before checking versions or compatibility.It invents a function, option, package, or file that is not in the evidence.It changes several unrelated files in one proposed fix.It gives a fix without stating the suspected root cause.It ignores a caused-by exception, assertion difference, or exit code.It claims success without a verification command or observable result.

RISK SCORE 0/19

No obvious warning signs

Still verify the diagnosis against the source code and reproduce the result.

AI debugging Bingo

Mark a square whenever your assistant reaches for a cliché

0/9 marked · saved in this browser

Have a better AI failure story?

Share what the assistant said, what the actual problem was, and the question that finally produced a useful answer. Public submissions live in the community issue tracker so others can reply.

Share a field note