The blog post "Stop Telling Me to Ask an LLM" reached 175 points and 102 comments on Hacker News after flagging on the platform.
It argues that defaulting to large language models wastes time and introduces unnecessary error for tasks better solved by direct lookup, scripts, or existing tools.
The Core Argument
The post lists repeated advice that sends users to LLMs for factual retrieval, simple calculations, or API calls. These cases produce hallucinations or extra latency when a one-line command or browser search returns the answer in under 200 ms.
HN commenters documented the same pattern across code debugging, unit conversion, and documentation lookup.
When LLMs Add Friction
Default LLM use creates measurable overhead. Average prompt round-trip on consumer hardware is 800-1500 ms. Direct shell commands or cached database queries finish in 5-50 ms.
Error rates rise when models fabricate version numbers or package flags that no longer exist. The post cites cases where users spent minutes verifying LLM output that a single man page or curl request would have settled instantly.
Decision Framework
Apply this three-question check before prompting:
- Is the answer already in local docs, man pages, or a stable API?
- Can a 10-line script or one-line pipeline produce it deterministically?
- Does the task require synthesis across sources that change daily?
If any answer is yes, skip the model.
Comparison of Approaches
| Task | LLM Path | Direct Tool | Time Delta |
|---|---|---|---|
| Package flag lookup | 1.2 s + verification |
man or --help
|
0.05 s |
| JSON schema validation | Prompt + parse |
jq or ajv
|
0.01 s |
| Unit conversion | Model arithmetic |
units or numfmt
|
0.02 s |
| Latest release check | Model cutoff risk | GitHub API | 0.3 s |
Direct tools win on speed and determinism for these operations.
Who Should Apply This Rule
Developers maintaining scripts or internal tools gain the largest time savings. Researchers verifying citations should route first to original sources rather than models.
Teams building retrieval-augmented systems can reduce token spend by 15-30 % by adding a pre-filter that routes deterministic queries away from the LLM.
Skip the rule only when the task genuinely needs multi-hop reasoning or creative generation.
How to Implement the Filter
Add a lightweight router before the main model call. Check query type against a short list of patterns (regex for commands, known API endpoints, local file paths). Route matches to the appropriate tool and return results without invoking the LLM.
Existing libraries such as LangChain's routing chains or simple if-else guards in Python achieve this in under 50 lines.
Bottom line: Treat "ask an LLM" as the last resort, not the default, for any query that existing deterministic tools can answer faster and more reliably.
The pattern will spread as teams measure actual latency and error costs rather than defaulting to the newest model.
Top comments (0)