AI is a tool and using it well takes skill

TL;DR Highlight

A rebuttal to Hickey: LLMs are tools, and skill in directing and checking them decides the outcome

Who Should Read

A developer who uses LLMs for coding and analysis and is unsure how much to delegate versus verify

Core Mechanics

LLMs can be wielded like tools
- The author argues that even under Rich Hickey's own definition, an LLM can be something a human wields.
- Like a saw not guaranteeing a straight cut, results depend on the operator's judgment.
Unpredictability is a flaw, not disqualification
- Predictability, reliability and simplicity are criteria for judging tools, not requirements for being one.
- A tool stays a tool even when it is hard to control or easy to misuse.
Automation and tools are not opposites
- Compilers and spreadsheets do a lot of work without step-by-step direction.
- We still wield them by specifying inputs, checking results and deciding next steps.
LLM use has a feedback loop
- A programmer can have code explained, challenge it, inspect the source, test a change, reject it and refine the request.
- Knowing what to ask, what evidence matters and when an answer is wrong makes the tool more useful.

Community Reactions

Opinion that less skill is needed as models improve
- One commenter said workflow tweaks were needed in mid-2025 but context management and special recipes now matter little.
- Another said a one-line prompt to root-cause a log got a correct answer in about 10 minutes, sometimes needing one steering follow-up.
Counterpoint that skill still matters
- One commenter said that at every model iteration the 'it still takes skill' group turned out right in hindsight.
- They said Opus 5.5 still makes mistakes all the time.
Experience that unguarded AI causes complexity creep
- A commenter said complexity quickly becomes unmanageable for humans and leaves you dependent on AI.
- They said even Claude runs away on larger projects, so professionals must contain complexity.
Opinion that it feels more like a slot machine
- A commenter said tools have repeatable cause and effect, while LLM results sometimes come only by luck.
- Another reported losing touch with the systems they manage when using AI.

How to Apply

Add a rebuttal follow-up to every LLM answer
- Today, after getting a code explanation or incident analysis, send 'where could this be wrong and what evidence can I check?'.
- Verify the evidence yourself in the source or logs.
Make an AI-output verification step explicit
- Add 'Was the AI suggestion verified by tests?' to your code review checklist.
- Set a team rule not to accept AI output until tests run or the source is checked.
Read first, then use the LLM to question
- When studying a paper or docs, read it yourself and write your interpretation.
- Paste the interpretation into an LLM and ask for objections and areas to examine.

Try It

copy & run

# Paste into Claude or ChatGPT: simplified version of the article's feedback-loop idea as a prompt

Explain the code below.
{{code}}

After explaining, answer the following.
1. Three points where your explanation could be wrong
2. How I can verify each point myself (source location, tests, logs)
3. If you propose a change, what it could break

Next step: actually run the verification methods it gives, and if something is wrong, paste the result and ask again.

Terminology

Rich HickeyCreator of the Clojure language. This post rebuts his article 'AI is Not a Tool'.

WieldTo hold and direct a tool. The user sets direction, sees results and corrects.

Anthropomorphic vocabularyPhrases like 'go rogue' that attach human intent to software. The author says they suggest human motives without establishing any.

Feedback loopThe cycle of asking, verifying and re-requesting. It is where the user's judgment enters.