A thesis with AI agents: template and citation rules

Final semester has the same pattern every year: the proposal is waiting, chapters are piling up, and AI agents get used to help, sometimes in a safe way, sometimes in a way that gets you called in by your supervisor. The difference comes down to one thing: clear rules of the game before starting.

This article summarizes the thesis-agent template we prepared: the standard chapter structure, citation rules that put honesty above laziness, and how to control the cost so thesis experiments don’t turn into a balance leak.

First: what is and isn’t allowed

AI agents (Claude Code, Cline, and the like) are used in two modes:

Allowed and useful:

Explaining concepts and theory until you truly understand them.
Building the chapter and sub-chapter outline from a framework you prepared.
Helping type the draft with sources THAT YOU supplied.
Fixing grammar and citation formatting.
Creating and running experiment code whose results you verify.

Not allowed:

Inventing sources: paper titles, author names, DOIs, years that don’t exist.
Inventing or “smoothing over” research result data.
Writing whole chapters whose content you don’t understand.
Helping with anything your campus guidelines forbid, in any form.

Rule number one of this kit: every factual claim must have a source a human can verify. If you’re not sure the source exists, the claim gets deleted or flagged, not kept. The full version is at the bottom of this article.

The standard chapter structure (and the “campus wins” rule)

The most common structure in Indonesian theses:

Chapter I (Introduction): background, problem statement, scope, objectives, benefits, systematics.
Chapter II (Literature Review): related research, theoretical foundation, framework.
Chapter III (Methodology): research type, data, instruments, analysis techniques, or a development method + design for build-and-design theses.
Chapter IV (Results and Discussion): findings, tables accompanied by narrative, comparison with the Chapter II theory.
Chapter V (Closing): conclusions that answer the problem statement, suggestions that connect back to the scope.

Three notes that often save a thesis: one paragraph, one idea; a table without narrative counts as unfinished; and a Chapter II never cited again in Chapters III/IV usually needs cutting.

The rule more important than all of them: if your campus guidelines differ, your campus guidelines win. This template is a general default, not a replacement for the guidelines.

Citation rules for working with AI

The most common violation pattern isn’t classic plagiarism but fabricated citations: sentences that sound right, with sources that never existed. AI can misremember a paper title, merge two authors, or fill in a year with a plausible-looking guess.

The practical rules of this kit:

The agent may not invent titles, names, DOIs, or years. If it isn’t sure the source exists, the agent writes [NEEDS SOURCE: ...] and stops there.
You supply the sources for the draft: PDFs in a repo folder, or links that open.
Every sub-chapter draft goes through a citation check: claims have citations, citations exist in the bibliography, nothing is swapped, and you can show the file for every source.
Paraphrasing still needs a citation. Copying a sentence needs quotation marks.

The clause is strict: claims without a defensible source get deleted, not kept “for now, we’ll find it later”. The full rules are at the bottom of this article.

Cost control: budget limits per key

An agent that runs continuously needs an API key, and this is where students usually get caught between two horror stories: the credit card they don’t have, or a balance draining by itself with no idea what it was spent on. Both are solved the same way:

Sign up and top up via QRIS starting at Rp10,000; no credit card in the flow.
Set a per-key budget limit; the default we recommend for a thesis is Rp50,000. Usage stops at the limit; the balance cannot go negative. Even the most wasteful task stops at the number you set.
Pick the model for the job: a flash model for drafts and cleanup (the jobs that run most often, the cheapest), a big model only for the parts that need heavy reasoning.
Per-request reports: you can see each agent session’s cost. This is practical when you split costs with your group: the numbers are open, not a matter of opinion.

The full setup how-to is in paying for AI APIs in rupiah with QRIS; the flow is the same as a professional developer’s, just with a smaller amount.

Three habits that make the result defensible

Work per sub-chapter, review before moving on. “Write the whole chapter” produces text you don’t master; sub-chapters you review one by one produce writing you can defend in the exam.
Keep an agent usage log. Date, sub-chapter, model, cost. If your examiner asks about the process, this log is the honest and orderly answer.
Check citations on every draft, before moving on. After all chapters are done, citation checking becomes a giant job; per sub-chapter it’s a two-minute habit.

Summary

AI agents speed up the parts that can be sped up: outlines, drafts from sources you supplied, experiment code, grammar. Citation honesty and understanding the content remain your responsibility, and the cost can be locked with per-key budget limits starting at Rp10,000.

The full template (agent rules, chapter structure, citation rules) will be published publicly ahead of the February proposal season; meanwhile, the pattern in this article is enough to start correctly.

Start with Rp10,000 →

Appendix: short rules for the agent (paste into your thesis repo)

Paste this as AGENTS.md in your thesis repo so the agent reads it before working:

- You are a research assistant, not a thesis writer. The student is
  responsible.
- Factual claims must have a source that can be verified. Not sure the source
  exists? Write [NEEDS SOURCE: ...]. Sources must not be invented.
- Research result data only from real result files in the repo.
- Work per sub-chapter; the student reviews before moving on.
- API keys never go into files/commits; use environment variables.
- Default per-key budget limit: Rp50,000. Raise it only with the student's
  explicit approval, recorded in the log.
- Campus guidelines always win over the rules in this file.