Run a long thesis project
Months of iterative revisions on a thesis, where the file, not the conversation, is the only thing that remembers what state you're in.
Copy this prompt to start your first task:
What is this workflow?
"Run a long thesis project" is an ongoing assistant relationship for months of iterative thesis revisions, where the connected draft file — not the chat history — is the single source of truth for what state you're in. Every real revision ships as a new versioned .docx with a short changelog of what changed and what's worth doing next, never a silent overwrite of the file you were already working from. Citations always carry a real URL, DOI, or journal/conference name, or an explicit "unconfirmed" flag instead of a number that looks real but isn't. Nothing ever gets deleted — only appended — so you always have a full version trail to roll back through.
This keeps your thesis moving between sessions. It doesn't stand in for your advisor or your own academic judgment. It won't email your advisor or a journal on your behalf, and while it'll give you an honest opinion on whether an argument holds up, whether your advisor will actually accept it is your call to make, not its. It never invents a citation, a data point, or a source: "unconfirmed" is the answer when it can't verify something, not a plausible-looking fake. And it never deletes an old version to make room for a new one; every revision is additive, so you can always tell it "revert to 0.9" and mean it.
How to use this workflow
This runs on one discipline: the connected file, not the chat, is the source of truth. This flow also never reaches outside Helio on its own — it won't email your advisor or a journal for you — so once it's set up, the routine work of drafting, versioning, and updating the file runs on its own without pausing for approval. See Control for what does pause for sign-off in general, like an action that does reach outside Helio.
Hire an Deep Research Writer, then rewrite the bio for file discipline, not the generic lab-teammate framing
Open AI Teammates in the sidebar, click Hire AI Teammate, and pick the Deep Research Writer template. Its default bio is framed around a lab research team (experiments, contribution claims, positioning against prior work). Replace it on the Profile step so it starts out already running your actual SOP: file as the only real state, version-forward not overwrite, transparent sourcing, and flagging its own contradictions unprompted:
Confirm the name and click Create teammate. They land in their own channel, which is where the whole thesis workflow below happens. More on hiring →
Connect the shared folder: the actual project home
The project folder holds the current .docx, every historical version, and an appendix .md for data. Connect wherever that folder actually lives (Google Drive, Notion, or a synced local path) so the teammate can read and write there directly instead of working from pasted-in fragments. Connect your tools →
One Google Drive caveat: Helio's Drive connection runs on the narrow drive.file scope, and its visibility is limited to files the teammate itself created or uploaded — there is currently no way to point it at a folder that already exists in your Drive (no file picker, no share-with-the-teammate path). So for an existing thesis project, either keep the folder in Notion or a synced local path, or attach the current .docx and appendix in chat once and let the teammate re-home them in Drive — from then on Drive works as its delivery surface and every version it writes there is fully visible to it. Since the deliverable is a .docx, also install the document-handling skill bundle (Skills & Plugins → the pdf/docx/xlsx/pptx bundle) so the teammate can actually create and edit Word files.
Every real edit produces a new version: nothing gets overwritten
This is the core discipline the SOP runs on: minor revisions step the version number forward (0.11 → 0.12 → 0.13), milestone-ready drafts jump to 1.0 / 2.0, and the old file always stays on disk as a fallback. The worked examples from here on come from a real Chinese-language thesis project, so the copy-ready prompts are in Chinese — the discipline they demonstrate is language-independent, so swap in your own thesis's language when you use them. Watch it play out on a routine edit:
Terse commands work: full sentences aren't required
Because progress lives in the file, not the chat history, you don't have to re-explain context every time. Two-word commands like "继续" (continue), "出docx" (give me the docx), or "3.2节那里" (that part in 3.2) work because the teammate reads the file's actual state to figure out what you mean, not because it's guessing or needs a paragraph of setup first: something a model without file discipline can't do safely.
A claim without a real source gets flagged, not invented
Data-source transparency is non-negotiable in the SOP: "待确认" beats a fake number. Ask it to add a claim that doesn't have a checkable source yet, and it says so instead of manufacturing a citation that looks real:
Turn on the daily research digest (optional, and it won't nag)
Open Automations in the sidebar and click Create via Chat. On Home, choose this teammate and send the prefilled scheduled-task starter. When the interview asks what to schedule, use the recurring job below. The digest is a one-way report: new findings, what they mean for a specific chapter or argument, and the update status. The teammate won't chase you for a reply if you don't send one:
Tips for Better Results
| Tip | Description |
|---|---|
| Connect the real project folder, not pasted fragments | The teammate reads and writes directly in Google Drive, Notion, or a synced local path. For an existing Drive project, attach the current .docx and appendix once so the teammate can re-home them — Helio's Drive scope can only see files it created or uploaded itself. |
| Let terse commands do the work | Because state lives in the file, not the chat, short commands like "继续" (continue), "出docx" (give me the docx), or "3.2节那里" (that part in 3.2) work without re-explaining context. |
| Give it a real source, or expect "unconfirmed" | Any citation without a checkable URL, DOI, or journal/conference name gets flagged "unconfirmed" instead of invented. Bring the original source if you want it fully cited. |
| Use version numbers to talk about state | Minor edits step forward (0.11→0.12→0.13); milestone-ready drafts jump to 1.0/2.0. Say "revert to 0.9" and it's a real rollback, since nothing is ever deleted. |
Common Questions
Industry & Specialized
Back to the Industry & Specialized use case overview.
Source and screen candidates
An HR teammate that finds passive candidates and screens against a real bar.
Get a personal AI assistant
One teammate for whatever comes up.
Source and screen candidates
Push back on the vague JD, go find people who aren't applying, and reach out to each one like you actually read their background, not a keyword-matched resume screen.
Run your company solo with AI
Hire across marketing, support, and ops (the functions a solo founder can't cover alone) and put them in one channel so the company runs without a headcount.