Using AI Efficiently: Tokens, Files, Workflows

1. Purpose of this guide

Many students use AI inefficiently: they upload entire PDFs, ask very general questions and get long answers back that they can barely use. This uses up limits and often leads to worse answers.

Efficient AI use does not mean asking as little as possible. It means: asking the right question with the right material in the right tool.

2. Understanding usage limits and tokens

AI systems do not work with pages, but with small text units. These are called tokens. A token can be a short word, part of a word, a punctuation mark or a fragment of a longer word.

What matters most for you: everything the AI has to read or write uses up capacity. This includes your question, uploaded files, earlier chat messages, the AI's answer and sometimes internal intermediate steps or tool work.

Today's models often have very large context windows, in some cases up to a million tokens. However, this does not mean the model weighs the entire content equally: even with a large context window, it pays to keep material short and upload only what is relevant.

What does this mean in practice?

A 25-page PDF does not cost "25 pages". It costs whatever the extracted text, formatting remnants, tables, footnotes and your question cost together. If the AI then writes a long summary from it, that output uses up capacity again too.

Stronger models do not automatically use more tokens per word. But they often have tighter usage limits and, for complex tasks, involve more internal processing. This is why you should save strong models for difficult reasoning tasks: checking arguments, comparing theories, analysing unclear passages. For mechanical work such as conversion, sorting or an initial rough clean-up, a simpler tool is often enough.

3. What wastes tokens unnecessarily

Most limits are used up not by good academic questions, but by untidy inputs.

4. The main rules for saving tokens

These rules do not just help with limits. They also improve accuracy. The less irrelevant material sits in the context, the more easily the AI finds the parts that matter.

5. Using models strategically

Not every task needs the strongest model. If you do everything with the most expensive or most tightly limited model, you use up your limits on routine work.

First lever: adjusting the reasoning effort

Before you even switch models, a simpler lever is worth trying: reasoning effort. Paid ChatGPT plans let you set an effort level, Gemini offers "Fast" and "Thinking" modes, and Claude adjusts its own reasoning effort to the task. Low effort is enough for mechanical work such as converting, sorting or cleaning up formatting. High effort is worthwhile when you want to check an argument or understand a difficult passage.

An up-to-date overview of the models by provider and plan is in the model comparison table in the AI Guide.

Splitting up the workflow

If you are allowed to use several permitted AI systems, you can split up the work. For example: one tool converts a PDF into Markdown; another tool reads the clean Markdown file and helps with the analysis. This way you do not spend all your limits on preparatory work in a single system.

This is not a trick for getting around the rules. It is normal work organisation: mechanical preparatory work and intellectual analysis are kept separate.

6. Writing prompts that cost less

An efficient prompt is not long. It is bounded. It tells the AI exactly what to do, with which material, and how short the answer should be.

Minimal efficiency prompt

Use only the following section.
Task: [specific task]
Answer in a maximum of [number] bullet points.
If the information is not in the section, write: "Not stated in the section."

Example: understanding a methods section

Use only the following methods section.
Explain in a maximum of 6 bullet points:
1. Which data are used?
2. How is it analysed?
3. Which limitation of the method does the text itself name?
No general summary.

Example: reading a theory section in a targeted way

Use only this theory section.
Which definition of [term] is used here?
First give a one-sentence answer.
Then a maximum of 4 bullet points with justification from the text.

Why this saves

You limit the input, the task and the output. The AI does not need to explain everything, only answer the question. This makes the answer shorter and usually more accurate.

7. Preparing files in an AI-friendly way

PDFs are made for people. For AI they are often untidy: columns, footnotes, headers, page numbers and hyphenated words are sometimes read incorrectly.

Current models mostly read PDFs directly, including layout and figures. Converting to Markdown or TXT therefore mainly saves capacity and gives you control over the material; it does not automatically make the answer more accurate. For linguistics work: keep the PDF when examples, glosses, IPA characters or syntax trees matter, because these are exactly the elements that often break during conversion.

Intro-
duction        3

THE THE-
ORY OF...

Sluckin 2025 | page 4

# Introduction

The theory of ...

Which formats are useful?

• PDF: original, page numbers, quotations, figures, final check.
• Markdown (.md): structured working version with headings, lists, examples.
• TXT: plain text, especially economical, good for short sections.
• DOCX: good for submission and correction processes, but not always optimal for AI evaluation.

The working version never replaces the original. It is only the AI-friendly version for working with.

8. File and code tools in the chat: convenient, but not limit-free

Some AI systems can process files in a working environment inside the chat. There you can upload a file and have it converted into Markdown or TXT. For many students this is the simplest way.

ChatGPT, Gemini and Claude can all run code or generate files in some form. With Claude, file generation is part of every plan, including the free one.

When are these tools useful?

• when you do not want to use a command line;
• when a PDF has many broken line breaks;
• when you want to merge several files;
• when you want to turn a Word or LaTeX file into a working version;
• when tables or examples need to be flagged for you to check later.

Conversion prompt

Convert the uploaded file into clean Markdown.
The goal is a working version for later AI evaluation, not a nice layout.

Keep:
- headings
- paragraphs
- numbered examples
- tables, where sensible
- existing bibliographic references

Remove:
- headers
- footers
- page numbers
- incorrect line breaks
- hyphens from hyphenation

Do not invent anything. Mark unclear spots with [check in original].

For LaTeX files

Convert this .tex file into Markdown.
Remove technical LaTeX structure.
But keep the academic content:
- headings
- body text
- examples
- tables
- citation keys
- important formulas or structures

If something cannot be converted cleanly, mark it with [check in original].

9. Pandoc: converting locally, saving AI limits

Pandoc is a program that can convert documents between formats. It is particularly useful for Word, Markdown, LaTeX, HTML and TXT. The advantage is simple: the conversion runs on your computer and does not use up any AI limits.

Installing on Windows

1. Search online for "Pandoc download" or go directly to pandoc.org/installing.html.
2. Download the Windows installer. The file usually ends in .msi.
3. Install Pandoc.
4. Open PowerShell or Command Prompt.
5. Check with:

pandoc --version

Installing on Mac

Simplest way: search online for "Pandoc download" or go directly to pandoc.org/installing.html, download the macOS installer package and install it. Then open Terminal and check:

pandoc --version

If Homebrew is already installed, this also works:

brew install pandoc

Key commands

# Word to Markdown
pandoc input.docx -t markdown -o output.md

# LaTeX to Markdown
pandoc input.tex -t markdown -o output.md

# Markdown to Word
pandoc input.md -o output.docx

# Markdown to TXT
pandoc input.md -t plain -o output.txt

# merge several Markdown files
pandoc chapter1.md chapter2.md chapter3.md -o working-version.md

Pandoc is often not ideal for PDFs, because PDFs do not have to contain a clean text structure. For PDFs, the file and code tools in the chat or a good PDF export tool are sometimes more practical.

10. Projects and knowledge collections

Many AI systems offer project areas, spaces or knowledge collections. There you can store files and instructions relating to a topic.

This is efficient because you do not have to explain your research question, outline and sources again in every chat. But it is not a magic memory.

How such systems usually work

The system does not always read all the files in full for every answer. It often searches your project files for matching passages and uses those for the current question. This saves context but can miss relevant passages.

Good project structure

term-paper/
├── 00_research_question.md
├── 01_outline.md
├── 02_sources.md
├── 03_excerpts/
│   ├── source_1.md
│   └── source_2.md
├── 04_key_examples.md
└── 05_open_questions.md

Project prompt

For this question, use only the files 00_research_question.md and source_1.md.
Answer the question concisely.
If the information is not there, write: "Not stated in the project material."

Why Markdown/TXT helps especially here

Clean, short files are easier to find than twenty uncommented PDFs. Every file should briefly state at the top what it contains.

# Excerpt: Müller 2020
Topic: language change and variation
Relevant for: chapter 2 of my term paper
Status: checked against the original; page numbers still to add

Memory

ChatGPT, Gemini and Claude can all remember content across individual chats: ChatGPT and Claude each have their own memory feature, which you can view, edit or delete, while Gemini bundles personal settings under "Personal Intelligence". For work on your thesis, it is better to use project-only memory (ChatGPT: project-only memory; Claude: project-only memory) or a temporary or incognito chat (ChatGPT: "Temporary Chat"; Claude: incognito chat), so that nothing carries over into other chats. Do not store drafts or interview or informant data in the memory of an AI system.

11. Agents and browser extensions

AI providers offer agent features that can carry out several steps independently or act in the browser, such as "ChatGPT Work", "Claude in Chrome" or "Claude Cowork". With ChatGPT and Claude, these features are only available on paid plans.

12. Three efficient workflows

Use only the abstract, introduction and conclusion.
Give me a maximum of 7 bullet points:
1. Research question
2. Method
3. Result
4. one limitation of the study
No detailed summary.

Compare only source_1.md and source_2.md.
Criterion: how do both texts define [term]?
Answer in a table with a maximum of 5 rows.
No general summary.

13. Checking after conversion

A cleaned-up Markdown or TXT file is only useful if it is correct. So check it briefly before you use it further.

For linguistics work in particular: do not adopt examples, glosses and trees blindly. If something looks broken, check the original.

14. Summary