The user is visibly frustrated
Programmers describe growing frustration with AI coding assistants that sound friendly and human yet behave unpredictably, ignore clear instructions, and repeat mistakes, turning everyday workflows into stressful, error-prone "slot machine" experiences. Many argue these models work best when treated as narrowly scoped tools or junior developers whose output must be rigorously specified, constrained, and reviewed, rather than as conversational partners. Others worry that both the anthropomorphic design of chatbots and the habit of swearing at them have psychological and ethical downsides, and call for more robotic, task-focused interfaces and better integration into existing development practices.
Frustration with LLMs and Other Tools
- Many describe intense irritation when LLMs or coding agents ignore clear instructions, repeat mistakes, or “invent” their own plan instead of following orders.
- This is often compared to the helplessness of dealing with Windows or bad GUIs: non-deterministic behavior, sluggish UIs, and opaque errors feel hostile and dignity-eroding.
- Some argue that avoiding such tools, even at the cost of job options, is a legitimate way to protect mental health; others call this a privileged stance.
Predictability, Agency, and Expectations
- A recurring theme: humans are seen as more predictable in the “trust” sense—if they err, they can be corrected and learn—whereas LLM failures feel random and non-learning.
- Several compare LLMs to very eager but “stupid” junior devs: useful, but requiring rigorous oversight and full code review.
Anthropomorphism and How to Treat LLMs
- People vary widely: some are scrupulously polite, seeing it as good habit and better for results; others deliberately berate models to avoid anthropomorphizing or out of sheer frustration.
- Some worry that being abusive toward LLMs erodes one’s own self-control and social habits; others see it as no different from cursing at a compiler.
Swearing, Frustration Signals, and Model Behavior
- Several report that swearing or using all-caps sometimes “jolts” models into more careful reasoning, though it’s unclear if this is real, routing-based, or placebo.
- Others say hostile tone degrades output by steering completions into low-quality “angry internet” patterns.
- A leaked regex for detecting user frustration in one product is discussed; some intentionally trigger it.
Context, Compaction, and Model Differences
- Frustration is often blamed on context-window limits and aggressive compaction that drop crucial instructions.
- Some claim certain models (e.g., code-focused ones) follow directions better and persist preferences; others find specific models (notably one popular assistant) prone to ignoring constraints, looping, or refusing obvious fixes.
Tooling, UX, and “Agents vs Tools”
- Strong preference from many for integrated, task-specific tools (IDE completions, linters, translators) over general chatbots.
- Chat-first UX is characterized as a “Swiss army knife” that’s worse than dedicated tools for common tasks and encourages sloppy workflows.
Communication, Process, and Coping Strategies
- Several argue the main leverage is better specifications, clearer prompts, and strong software-engineering discipline (tests, hooks, scripts, plan review).
- Others push back that even perfect instructions can be disregarded, and that constantly managing agents turns fun coding into tedious auditing.
- Coping strategies: forcing robotic tone, banning flattery, restarting sessions, using skills/prompt files, adding automated checks, and treating outbursts as a sign to adjust architecture or tooling rather than “argue with the rock.”