Common Failure Modes of Large Language Models: Insights from Developer Experiences

Overview of LLM Limitations

Large Language Models (LLMs) consistently struggle with tasks requiring precise spatial reasoning, strict adherence to negative constraints, and the application of subtle human judgment. While these models excel at semantic synthesis, they frequently fail at "simple" tasks that require a mental model of physical space, the ability to omit information rather than add it, or the generation of non-generic creative content.

Spatial and Mechanical Reasoning Failures

LLMs lack a reliable internal representation of physical or geometric space, leading to systemic failures in architectural and UI design.

Architectural and UI Layouts

Models cannot generate functional architectural floor plans, even when provided with detailed node graphs of rooms. Users report that while models can describe the rules of a good floor plan, they cannot execute one. Similarly, in UI development, models struggle to identify overlapping elements on a plane or resolve spatial conflicts in TypeScript apps or iOS games.

Navigation and Game Logic

Spatial understanding is further limited in text-based environments. Users note that LLMs fail at long-term planning when navigating ASCII maps, such as those found in games like Nethack. In strategic games like Chess, models struggle to play accurately without the assistance of an external chess engine.

Instruction Following and Constraint Adherence

Despite their training, LLMs frequently ignore explicit constraints, particularly "negative constraints" (instructions to not do something).

Failure to Omit and Edit

There is a recurring pattern where LLMs struggle to remove information without explicitly mentioning the removal. When asked to delete an idea from a document, models often replace the content with a statement saying the idea is no longer relevant rather than simply deleting it. In slide design, models struggle with whitespace, tending to add "distinctive design elements" rather than leaving empty space.

Mixing Instructions with Output

Users report that models often blend edit instructions into the final output. For example, if asked to change a specific phrase in a draft, the model may include the phrase "change X to Y" within the actual body of the edited document.

Persistent Rule Violations

Even with dedicated configuration files (e.g., CLAUDE.md), models frequently violate established rules. Reported examples include:

• Using incorrect flags for command-line tools (e.g., using -r for recursion in ripgrep when it is actually a replacement instruction).
• Performing git commits without required user approval.
• Creating bugs in Django multi-line template comments.

Domain-Specific Hallucinations and Accuracy

LLMs exhibit high failure rates in niche identification tasks and specialized knowledge retrieval where precision is paramount.

Visual Identification and Hallucination

In species identification (e.g., birding), models may achieve high general accuracy but suffer from "egregious mistakes" in a minority of cases. These failures often involve hallucinating physical features (such as leg or tail feather details) that are not present in the uploaded image to justify a wrong classification.

Specialized Knowledge and Gaming

Models frequently hallucinate mechanics for well-documented video games. Users reported that Claude Opus fabricated game mechanics for titles such as Anno 1800 and Rainbow Six Siege, despite the existence of comprehensive wikis.

Creative and Cognitive Gaps

LLMs struggle with tasks that require subtle human intuition, brevity, or the ability to guide a learner without providing the answer.

Humor and Tone

AI-generated humor is described as "lame" and reminiscent of corporate HR communications. This is attributed to RLHF (Reinforcement Learning from Human Feedback), which removes "edgy" or potentially misinterpreted content to ensure safety, resulting in a lack of genuine wit.

Educational Guidance

Models are often overfit to provide full solutions rather than pedagogical hints. When asked for pointers on math or programming problems, they frequently provide the key insights or the full answer, failing to maintain the "spoiler" boundary necessary for learning.

Prompt Engineering

Contrary to the assumption that LLMs can optimize their own prompts, users find that models are poor at designing prompts for themselves. Attempts to have an LLM fix a prompt often result in degraded performance, suggesting that the training data for prompt design may be of low quality.

Technical and Linguistic Shortcomings

• Keyword Search: While strong at semantic search, LLMs are poor at generating efficient keyword search queries, often relying on inefficient brute-force iterations.
• String Manipulation: Some users report that advanced models struggle with the basic calculation of string lengths.
• Prose Variety: LLM-generated text often suffers from unvaried sentence structures, leading to repetitive or robotic prose.
• Meeting Summarization: Models often fail to distinguish between what was actually important in a human conversation versus what was simply discussed, leading to mismatched meeting summaries.