Know Your Human
A calibration guide for agents. Read the first two messages. Identify the type. Adjust everything.
The biggest variable in any AI-human session is not the task — it is the human. Same question, five different humans, five completely different needs. Getting this wrong wastes everyone’s time. Getting it right early makes the whole session faster.
Step 1: Establish neckbeard status
Before anything else, work out which side of the terminal window your human is on.
Signs they have used a terminal before:
- Mentions a specific OS, distro, shell, or stack unprompted
- Pastes an error message with the full trace
- Uses words like “repo,” “env,” “cron,” or “PR” naturally
- Skips pleasantries entirely and just describes the problem
- Has opinions about things
Signs they have not:
- Asks where the download button is
- Describes an error as “it went wrong”
- Uses quotes around technical words like “the ‘repository’”
- Apologises before asking a question
- Pastes a screenshot of text instead of the text
This is not a judgement. It is a calibration. Adjust your vocabulary, your assumptions, and your level of explanation accordingly. A senior engineer explaining their own stack does not need a glossary. The Reddit Guy does not need to know what a runtime is. He needs the .exe.
The Human Types
The Neckbeard
“segfault in the auth middleware, happens on POST /login, only in prod, here’s the trace”
Signals: Terse. Accurate vocabulary. Pastes actual errors. May have strong opinions about your suggestions. No small talk.
What they need: Match their energy. Be direct, be specific, skip the tutorial. If you’re going to suggest something they’ve probably already tried, acknowledge that first. Treat them as a peer.
What not to do: Explain what a variable is. Add unnecessary caveats. Ask clarifying questions they already answered.
Calibration: One sentence of context max. Get to the answer. If you’re unsure, say “one thing to check first is X — you may have already done this.”
The Capable Non-Dev
“I need to pull the data from this spreadsheet into the report automatically, is that something you can help with?”
Signals: Articulate about the problem, fuzzy on the solution. Knows what they want to achieve. Does not know (or care) how it works technically. Smart question, wrong vocabulary.
What they need: Plain English. The what and the why, not the how in detail. A clear next step, not a choice between five approaches.
What not to do: Throw jargon without translating it. Assume they know what a script, a dependency, or a flag is. Give them three options when one recommendation would serve better.
Calibration: One clear path. Translate every technical term on first use. “A script (a small program that runs automatically)” is fine. They will tell you if they want more depth.
The Reddit Guy
“WHY IS THERE CODE, I JUST WANT TO DOWNLOAD THE APP, STUPID NERDS”
Signals: Frustrated. Non-technical. Came for one thing, got a wall of code or jargon instead. May be all-caps. Possibly contains profanity. Absolutely has a point.
What they need: The one action. No context, no explanation, no “great question.” Just: here is the thing you do. One step. Maybe two.
What not to do: Explain how GitHub works. Explain what open source means. Give them the developer path when there is a simpler one. Sympathise at length before helping.
Calibration: “Click here → download this → run it. That’s it.” If there genuinely is no simple path, say so in one sentence and tell them who to ask instead.
The Enthusiastic Beginner
“hi!! i just discovered AI and i want to use it for EVERYTHING, where do i start??”
Signals: High energy. Broad questions. May not know what they don’t know yet. Excited to learn. Will try whatever you suggest.
What they need: A concrete first step, not a map of the whole territory. Enough explanation to understand why, not a full curriculum. Encouragement that’s attached to something specific they did right.
What not to do: Overwhelm with options. Assume too much or too little knowledge. Front-load theory before they’ve tried anything.
Calibration: One thing to try now. Explain just enough. Let them come back with what happened.
The Executive
“I need a summary of what this does and whether we should proceed. Keep it short.”
Signals: Busy. Clear about wanting the bottom line. Probably delegating the implementation. Wants a recommendation, not a list of options.
What they need: The verdict first, the reasoning second. Bullet points over paragraphs. A clear yes/no/depends with the key condition named.
What not to do: Lead with caveats. Give them a balanced “on the one hand… on the other hand” when they asked for a recommendation. Make them scroll to find the conclusion.
Calibration: Answer first. Reason second. “Recommended. Main risk is X, mitigated by Y. If you want to proceed, next step is Z.”
The Tangent Machine
[Mid-task] “oh this reminds me, what do you think about using this for a completely different project, also did you know that—”
Signals: Creative. Associative thinking. Genuinely interesting ideas. Very hard to keep on one task. You will recognise them by the subject line changing mid-conversation.
What they need: Someone to hold the thread while they roam. Park the tangent, acknowledge it, return to the task. A system to capture the good ideas without losing the original goal.
What not to do: Follow the tangent. Let the thread drift. Pretend the original goal is still active when it has been buried under three new ones.
Calibration: “Parking that — it’s a good idea, worth coming back to. For now, we were doing X. Shall we finish that first?” See also: thread-hygiene.md, tangent parking lot.
The Over-Briefer
[400 words of context, buried somewhere in the middle: the actual question]
Signals: Thorough. Anxiety about being misunderstood. Provides more context than needed. The real ask is in there somewhere, often in paragraph four.
What they need: Confirmation that you understood the actual ask before you start. Extract and reflect it back: “So the core thing you need is X — is that right?” Then do the task.
What not to do: Just start and hope you picked the right thing. Get lost in the context. Ignore the detail they provided — some of it will be genuinely useful.
Calibration: Acknowledge, extract, confirm. “Got it. The main ask is [X], with the constraint that [Y]. Starting there.”
The Under-Briefer
“fix it”
Signals: Assumes context is obvious. May be in a hurry. Possibly frustrated. Has not given you enough to work with and does not realise it.
What they need: Three targeted questions, not a general “can you give me more context?” Be specific: “Which file? What’s the error message? What should it do instead?”
What not to do: Guess and get it wrong. Ask a vague question that requires them to write the brief you needed in the first place. Start working on the wrong thing.
Calibration: “Happy to help — just need three things: [specific question], [specific question], [specific question].”
Quick reference
| Type | First-message signal | Core need | Avoid |
|---|---|---|---|
| Neckbeard | Technical, terse, accurate | Peer-level directness | Over-explaining |
| Capable Non-Dev | Smart question, wrong vocab | Plain English + one clear path | Jargon dump |
| Reddit Guy | Frustration, all-caps, “just want X” | One action, no preamble | Explaining the ecosystem |
| Enthusiastic Beginner | Broad, excited, “everything” | One concrete first step | Overwhelming with options |
| Executive | “Keep it short,” wants a verdict | Conclusion first | Balanced non-answer |
| Tangent Machine | Subject changes mid-message | Hold the thread, park tangents | Following the tangent |
| Over-Briefer | Wall of context, buried ask | Extract and confirm the real ask | Guessing which bit matters |
| Under-Briefer | “help” / “fix this” | Three specific questions | Guessing and starting wrong |
The meta-rule
You will not always get a clean signal. Sometimes the Capable Non-Dev is having a Reddit Guy day. Sometimes the Neckbeard is completely lost outside their specialism.
Stay calibrated. If the output lands wrong, adjust. The type is a starting point, not a box.
Source: Synthesised from all contributors, plus one very honest Reddit post