Ask HN: What hacks/tips do you use to make AI work better for you?
Programmers and power users trade concrete ways they make large language models more useful, from generating throwaway scripts, Docker and infra configs, and documentation to acting as coding copilots, brainstorming partners, translators, and even home-automation controllers. Many emphasize workflow tricks—good custom instructions, tightly scoped context, persona-style system prompts, local tooling, and treating the model like a junior colleague whose output must be reviewed. Others voice deep frustration, arguing that LLMs often hallucinate, struggle with niche or fast-changing frameworks, and mostly excel at boilerplate rather than hard problems, highlighting a growing divide over where AI genuinely boosts productivity versus wastes time.
Lightweight automation & scripting
- Many use LLMs for “non‑critical” glue: shell/Python/R scripts, GitHub Actions, small one‑off tools, Excel formulas, Dockerfiles, K8s YAML, HomeAssistant automations, etc.
- Value is highest where correctness is “good enough” and the alternative is never writing the script at all.
Developer workflows & coding assistance
- Popular patterns: inline completions in editors, code explanation, refactoring, generating boilerplate tests, SQL queries, plots, Pandas transformations, and documentation/docstrings.
- Some dump entire codebases or many chunks into a model instead of sophisticated RAG; others carefully restrict context to a few small files and refactor into <200‑line modules for better results.
- Opinions diverge sharply: some claim 5–10× productivity and share non‑trivial projects largely AI‑written; others say AI code editors are consistently poor and have stopped using them.
- Domain matters: strong results reported for TypeScript/React, Python, data work, “stable” APIs; very poor for C++, fast‑moving frameworks (e.g., Next.js), niche AOSP internals, distributed consensus, or bespoke systems.
Prompting, instructions, and personas
- Heavy use of custom instructions and system prompts: e.g., ultra‑concise mode, no disclaimers, never mentioning being an AI, or specific behavior when “!!” appears.
- Some create “characters” (e.g., shell‑only bot, terse senior dev, code‑dumper bot) or ask the model to be opinionated/“an asshole” to extract clearer views.
- Emphasis on providing concrete context (code, docs snippets, project structure) and iterating; several argue that learning to ask precise, scoped questions is a new core skill.
Non‑coding and tooling uses
- Workflows include iOS Shortcuts calling APIs, cross‑provider desktop clients, translation between languages, voice dictation via Whisper‑like tools, parsing PDFs into systems with human review, brainstorming architecture, project planning, life/health coaching, and “roasting” to expose blind spots.
Skepticism, limits, and organizational impact
- Some experienced developers say they “cannot get these things to do anything useful,” citing hallucinated APIs, outdated patterns, and debugging overhead exceeding any gain.
- Others counter that this reflects domain, expectations, and lack of shared transcripts; they treat LLMs like junior devs or Stack Overflow++.
- One ERP firm reports replacing most full‑time devs with consultant‑plus‑LLM tooling, raising margins; others predict similar shifts for CRUD‑style work.
- Several note fundamental limits: models can’t truly “think,” are hard to debug, and excel mainly at repetitive boilerplate rather than deeper business logic.