LLM tokens
The chunks of text models bill and reason over; context windows cap how much instruction, job description, and candidate material fit in one call, which shapes how recruiters package prompts and attachments.
Michal Juhas · Last reviewed May 2, 2026
Who this is for
Anyone pricing high-volume screening or outreach, or wondering why the model "ignored" part of a giant paste.
In practice
- Summarize before you attach: JD essentials, top five must-haves, and one paragraph of employer pitch beat raw exports.
- Structure beats volume: tables with clear headers help models fill structured output style fields.
- Watch automation loops: a Make scenario that re-sends full histories every row multiplies token spend fast.
Where it breaks
Optimizing only for token count can strip compliance-relevant detail. Always keep a human-readable source of truth outside the model thread for audits and disputes.
Rough mental model
| Input style | Typical outcome |
|---|---|
| Lean Markdown SOP | Predictable, cheap reruns |
| Full PDF dump | Noisy parse, higher cost |
| Chat thread archaeology | Important lines may truncate |
Related on this site
- Glossary: Markdown for AI, System instructions, Large language model
- Blog: How to use AI in recruiting
- Course: Starting with AI: the foundations in recruiting
- Membership: Become a member
Frequently asked questions
Why should recruiters care about tokens?
They drive cost, latency, and truncation. Pasting a twenty-page PDF plus ten threads can crowd out your actual instructions or get cut mid-document. Markdown for AI and curated excerpts usually win.
Are tokens the same as words?
Roughly correlated but not identical: short common words can be one token; rare words or code may split. Vendors show estimates in product UIs; treat them as directional, not exact accounting.
How does this tie to system instructions?
System and user content share the same budget. Heavy boilerplate leaves less room for candidate specifics. That is why teams move stable rules into system instructions and keep each task message short and structured.
What about images or resumes?
Multimodal inputs have their own limits and pricing. OCR’d resume text still counts as tokens. Decide what must be in-model versus what stays in your ATS for human review, especially around hallucination risk.
Does a bigger context window fix everything?
No. Models can lose focus in very long contexts, and automation still needs monitoring. Better retrieval and smaller trusted snippets often beat "send the whole drive".
Where can we learn more practically?
Read How to write better AI prompts, tighten your Markdown for AI packs, and rehearse packaging in a workshop before you wire high-volume workflow automation.