No coding required. Just the concepts that make AI tools actually work for you.
Before you can understand anything else about AI, you need to understand tokens. They're how AI models read, process, and charge for text.
A token is the smallest chunk of text an AI model processes. It's not exactly a word — it's more like a word fragment. Common words are often one token. Longer or uncommon words get split into multiple. Spaces and punctuation are tokens too.
1,000 tokens ≈ 750 words. A typical job description is ~300–600 tokens. A 1-hour interview transcript is roughly 8,000–15,000 tokens.
*Estimate based on Claude Sonnet input pricing (~$0.003/1K tokens). Actual cost varies by model and usage.
If you're processing 50 resumes through a Zapier automation, token efficiency = real cost savings. A bloated prompt that says the same thing in 200 words vs 40 words costs 5x more — and doesn't get better results.
The context window is the total amount of text an AI can "see" at one time. It's finite — and more stuff is competing for that space than you might think.
Before you type a single word, these things are already in there:
| Model | Context Window | Approx. Word Equivalent | You Use It Via… |
|---|---|---|---|
| Claude Sonnet 4.6 | 200K tokens | ~150,000 words | Claude.ai Zapier AI |
| GPT-4o | 128K tokens | ~96,000 words | ChatGPT Copilot |
| Gemini 1.5 Pro | 1M tokens | ~750,000 words | Gemini |
Just because you can paste 20 resumes doesn't mean you should. The model performs better on content near the top and bottom of the context. Burying key instructions in the middle of a giant paste = worse results.
The amount of context isn't the only thing that matters. What's in it matters just as much — and there are two ways it can go wrong.
As a conversation fills up, early instructions get diluted by newer content. The model may start "forgetting" things you told it 20 messages ago — not because it can't see them, but because they're buried under so much other text.
TA Example: You told Claude "only evaluate candidates with enterprise experience" at the start of a long session. By message 40, it starts recommending mid-market candidates anyway.
One piece of wrong or conflicting information contaminates the whole session. Even after you correct it, the original (wrong) answer is still sitting in the context history — and sometimes the model drifts back to it.
TA Example: Claude misidentifies a candidate's current employer. You correct it. Later in the same chat, it references the wrong employer again.
Imagine handing your boss a raw transcript of a 2-hour meeting where you debated a decision, went back and forth, rejected a few options, and finally landed on one. They skim it. They might latch onto the rejected option from page 2 instead of the final decision on page 10. That's exactly what's happening inside a long AI session.
AI models sometimes state things that are completely false — with complete confidence. This isn't a bug that will be "fixed." It's an inherent feature of how these models work.
AI models are trained to predict the next most likely token. They're optimized to produce plausible-sounding responses, not necessarily true ones. When a model doesn't know something, it doesn't say "I don't know" — it generates what a correct answer would likely look like.
Never ask AI to recall candidate details it wasn't given. Always provide the resume or transcript as context. AI is great at analyzing what you give it — it's terrible at retrieving facts from memory.
Paste the source material. Ask AI to work from it. The more you ground a prompt with real context, the less the model needs to fill in gaps — and the less it will hallucinate.
Prompting is a skill. The same AI will give you very different outputs depending on how you phrase your request. Here's a simple framework that works.
If you want the AI to write in a particular style or structure, paste an example of what "good" looks like. "Format it like this: [example]" is more reliable than a long description of what you want.
Most people think AI is just one thing. It's actually four layers — each built on top of the last. Understanding the stack demystifies why the same "AI" behaves so differently across tools.
Click any layer above to learn more.
The core AI, trained by companies like Anthropic, OpenAI, or Google on massive amounts of text. It takes in tokens and predicts what tokens should come next. The model itself is just a big math function — it has no memory, no UI, no tools. It only exists as an API.
Examples: Claude Sonnet 4.6, GPT-4o, Gemini 1.5 Pro, Llama 3
The raw capabilities the model exposes. Think of these as the verbs it knows how to do. Every higher-level feature you use is built from combinations of these:
The harness is the software that orchestrates the model. This is where the real magic (and real complexity) lives. It's what turns a raw text-prediction engine into something useful. The harness handles:
Same model. Different harness. Zapier's harness has different system prompts, different tools connected, different memory behavior. That's why the same model can feel totally different across products.
The interface you actually use — the UI, the login, the feature set, the pricing. The product is the harness + a UI wrapped around a model. Two products using the same model and similar harnesses can still feel very different because of UI decisions, what tools they've connected, and what system prompts they're using.
Examples: Claude.ai, ChatGPT, Zapier AI Agents, Microsoft Copilot, Gemini
You'll hear "agent" constantly right now. Here's what it actually means.
You ask → AI responds → done. Like a search engine. The model gets your message, generates a reply, stops.
Example: "Summarize this JD." → Claude summarizes it → you read it.
You give a goal → AI figures out the steps → takes actions → checks results → repeats until done. The model is in a loop, making decisions.
Example: "Screen all applicants for this role and flag the top 5." → Agent reads each resume, scores them, ranks them, returns a shortlist.
The more steps an agent takes without human review, the more a single early mistake can compound. In TA workflows, always design agents to surface their reasoning for anything that affects a candidate decision — don't let them run fully unattended on consequential calls.
| Product you use | Underlying model | Harness built by |
|---|---|---|
| Claude.ai | Claude Sonnet / Opus | Anthropic |
| ChatGPT | GPT-4o | OpenAI |
| Microsoft Copilot | GPT-4o (+ Bing search) | Microsoft |
| Zapier AI Agents | Claude, GPT-4o, or others | Zapier |
| Gemini | Gemini 1.5 Pro |
Every AI model has a "knowledge cutoff" — a date after which it wasn't trained on new data. For anything that happened after that date, the model is flying blind.
Asking AI to research a candidate's current employer, recent funding, or headcount is risky — that info may be stale or fabricated. Use AI to analyze; use LinkedIn, Crunchbase, or the company's site to verify facts.
If you need AI to reason about something recent (a job posting, a company update, a candidate's LinkedIn), paste the text directly into the prompt. The model can work with information you give it — it just can't retrieve it on its own.
Apply these right away — no technical knowledge required.
Long threads accumulate context rot. Open a new chat when you switch topics or candidates.
Resume analysis? Paste the resume. Transcript summary? Paste the transcript. Don't rely on AI memory.
Don't ask "is this a good candidate?" Ask whether specific must-haves are met. Vague prompts get vague answers.
Want a table? A bulleted list? Three sentences max? Say so. Format guidance dramatically improves usability.
AI-generated claims about companies, candidates, or recent events should be verified before you act on them.
In a long session, restate important rules near the bottom before asking. "Remember: we are only considering candidates with X."
If the output isn't right, tweak the prompt — don't just ask again. Small wording changes make a big difference.
Candidate data is sensitive. Know your company's guidelines on what's OK to paste into which tools. When in doubt, check with your People team.
Five quick questions. Click an answer to see if you've got it.