Ran out of tokens? Continue it.

Ran out of road here?
Take your project with you.

Tokens run out. Chats reset. The AI you were using hits its wall in the middle of a project — and that used to mean starting over. It doesn’t have to. Freeze the project’s rules in one small file, keep its exact stopping point in another, and almost any AI can pick the work up where it stopped. This page is the whole flow: pick the next AI, build the pack, make it prove the resume took.

1 · Pick your next AI

Every major AI’s continuation route, grouped by how it holds a project — below.

2 · Build the pack

spec.md (frozen rules) + state.md (the stopping point). Two minutes, in your browser.

3 · Prove the resume

Paste the resume prompt first, and make the new AI state the goal, the stopping point, and the next action back to you.

Coming from a split run? Its run sheet already exports a state.md for the run — bring it here for the destination AI’s route. New to the method? Start with one sentence instead.

1 · Pick your next AI

Native — the AI has a project home

These AIs have their own projects and memory. Set the home up once, put the pack in it, and the project persists between chats — no pasting at the start of every conversation.

Native

ChatGPT

  1. Create a Project for the work — Projects are ChatGPT's persistent home — and turn Memory on.
  2. Upload spec.md and state.md into the Project; the files ride along into every chat inside it.
  3. Open a chat in the Project, paste the resume prompt (built below), and check its three answers before it starts.

Projects + Memory are officially documented. Project file caps were not verifiable on OpenAI's own pages when this map was checked (October 2026) — treat any number quoted elsewhere as unsettled.

Open ChatGPT ↗
Native

Claude

  1. Create a Project and add spec.md and state.md to its project knowledge.
  2. Turn Memory on — Claude keeps it as visible topic files you can read and edit, and it can search your past chats.
  3. Start the project chat with the resume prompt and check its three answers.

One catch, from Anthropic's own help pages: exporting your Claude data can't be started from the iOS or Android app — use the web or desktop app for that.

Open Claude ↗
Native

Gemini

  1. In the web app — Gems can only be created there, not in the phone apps — make a Gem for the project and put the spec in its instructions.
  2. Keep Personal Context on, so what Gemini learns carries between chats.
  3. Open the Gem, hand it state.md, and run the resume prompt.

Google has announced Gems become “Skills” starting November 2026, with existing Gems converted automatically — the home survives the rename.

Open Gemini ↗
Native

Mistral — Le Chat / Vibe

  1. Create a Project in Le Chat and upload spec.md and state.md to it.
  2. Coding too? The same home works in Vibe, Mistral's coding agent — Projects and the file Library are shared.
  3. Start each work chat inside the Project with the resume prompt.
Open Mistral — Le Chat / Vibe ↗
Native

Perplexity

  1. Create a Project (they used to be called Spaces) for the work.
  2. Upload spec.md and state.md to the Project.
  3. Chat inside the Project — Memory carries your preferences between chats, and the resume prompt re-anchors it to the files.
Open Perplexity ↗
Native

Genspark

  1. Put the project in a Hub — Genspark's persistent home for a body of work.
  2. Its SecondBrain keeps a git-backed Markdown store (SB-Git): spec.md and state.md live there as plain Markdown it can read back and diff.
  3. Resume with the prompt below — here, the files genuinely are the memory.

Documented in Genspark's own release notes and product pages (October 2026).

Open Genspark ↗
Native-lite

Copilot

  1. Copilot saves preferences (Saved Memories) but has no project home of its own on a personal account — so the pack does the carrying.
  2. In each work conversation, upload spec.md and state.md: officially up to 20 files per conversation, 50 MB each, Markdown included.
  3. Paste the resume prompt and check the three answers before it starts.

The Notebooks you may see mentioned are a Microsoft 365 work/school feature — not part of the personal Copilot route.

Open Copilot ↗
Native — provisional

Grok

  1. Safest route: start a chat for the project and upload spec.md and state.md into it.
  2. Grok syncs chats across devices and takes uploads and connected sources — but Projects and memory specifics were not documented on xAI's own help pages when this map was checked, so its memory is reported, not verified.
  3. Run the resume prompt and verify its three answers — with Grok, check them twice.

Provisional: this card will firm up when xAI documents Projects/memory officially.

Open Grok ↗

Project-file — no memory, but its prompts are files

No memory feature at all — instead, the prompt itself is a file you own, and reopening the file reopens the project exactly.

Project-file

Google AI Studio

  1. Write the project into a prompt in AI Studio and save it — AI Studio stores prompts as files in the AI_Studio folder of your Google Drive.
  2. Bake spec.md's rules into that saved prompt; keep state.md beside it in the same Drive folder.
  3. Next session, reopen the saved prompt from Drive — it loads back exactly — paste the current state.md under it, and run the resume prompt.

AI Studio has no memory; the Drive files are the whole pathway. That also makes this the most predictable route on the page.

Open Google AI Studio ↗

Paste-only — nothing persists; the pack is the memory

Between chats, these keep nothing — no files, no memory. That's fine: paste the pack (or upload it, where allowed) at the start of every chat, and update state.md whenever you stop. The discipline is the feature.

Paste-only

DeepSeek

  1. Start a fresh chat. Paste the resume prompt, then spec.md, then state.md.
  2. Work. The moment you stop — quota, a closed tab, anything — update state.md first. The next chat will not remember you.

DeepSeek publishes no consumer help pages documenting project storage or memory, and consumer limits quoted elsewhere are unverified — the paste route needs no limit to work.

Open DeepSeek ↗
Paste-only

Qwen

  1. Start a fresh chat. Paste the resume prompt, then spec.md, then state.md.
  2. Work, and update state.md before you leave — nothing carries over on its own.

No consumer help pages from Alibaba document a project store for Qwen Chat; limits quoted elsewhere are unverified.

Open Qwen ↗
Paste-only

Kimi

  1. Start a fresh chat. Paste the resume prompt, then spec.md, then state.md.
  2. Work, and update state.md before you leave.

Secondary roundups call Kimi the roomiest of this class for very long pasted context — that's secondary reporting, not an official Moonshot consumer page — but it's why Kimi is the paste-only pick when a pack is huge.

Open Kimi ↗
Paste-only

Z.ai

  1. Start a fresh chat. Paste the resume prompt, then spec.md, then state.md.
  2. Work, and update state.md before you leave.

Z.ai's consumer surfaces were the least documented of this class when checked — nothing official describes project storage. Paste the pack; trust nothing to persist.

Open Z.ai ↗
Paste-only

Meta AI

  1. Start a fresh chat. Paste the resume prompt, then spec.md, then state.md.
  2. Work, and update state.md before you leave.

Treat anything you hand Meta AI as belonging to that one session: no official Meta help page documents a project memory, and secondary accounts describe its files as session-bound.

Open Meta AI ↗
Paste-only

Venice

  1. Venice is request-scoped by design: include the pack in each request — upload spec.md and state.md (25 MB per file, per Venice's own docs), or point it at a public URL where the files live.
  2. Paste the resume prompt with them, every time. There is no later chat that remembers.
Open Venice ↗

No pathway — never the project's home

These cannot hold a project at all. Use them for single, self-contained jobs — never as the place a project lives.

No pathway

Arena

Never paste a private pack into Arena. It is an evaluation platform, and its terms say prompts you enter may be selected and published for research. There is also no project or memory to resume from. Fine for a one-off experiment with throwaway text — never the project's home.

Open Arena ↗
No pathway — practically

Duck.ai

Duck.ai stores nothing on its servers: chat history lives only on your device (Safari deletes it after 7 days), and documents can't be uploaded at all. A long project has nowhere to live here. One-off questions only.

Open Duck.ai ↗

Pathway map checked October 2026 against each vendor’s own help pages and documentation. Where a vendor hasn’t documented a feature, the card says so — nothing here is guessed. These features change often, and that’s exactly why the pack matters: the pack works even when a pathway rots, because the pack is yours.

2 · Build the pack

spec.md is the frozen rules — write it once; don’t edit it mid-project. state.mdis where the work stands — update it every single time you stop. Fill in what’s easy and skip the rest: blank fields still produce a valid skeleton you can complete later, never an error.

spec.md ≈ 676 characters · state.md ≈ 481 characters. Keep the whole pack modest — roughly 50k tokens of Markdown or less is a good working ceiling (guidance, not a vendor limit).

Everything on this page runs in your browser. What you type is not sent anywhere, is not stored by this site, and is gone when you close the tab unless you download or copy it first. The pack is yours the moment it exists.

3 · Prove the resume took

The resume prompt makes the new AI state three things back to you before it works: (1) the project’s goal and acceptance criteria, from spec.md; (2) the exact stopping point, from state.md; (3) the single next action.

If any of the three is wrong or vague, the resume didn’t take — re-supply the pack before letting it work. A fluent summary is not a resume; the three answers are.

The pack is yours

Keep spec.md and state.md where youkeep things: a folder in your Google Drive (the store the most AIs can read back), OneDrive if Copilot is your main AI, iCloud Drive if you’re Apple-only, or a plain folder on your computer or tablet. No account here, no lock-in — two Markdown files, readable by everything, owned by you.

On Apple gear, honestly:an iPad can continue a prepared project — the files ride in iCloud Drive and you hand them to the AI app — but the setup is web-first (Gems are created on the web; Claude data export can’t start from the phone apps), and the last step is always a manual paste or upload. That final tap can’t be automated away; plan for it instead of fighting it.

🪫 Run a job across free models →✍️ Start a new project🤖 All the AIs, one list

The pathway map comes from an October 2026 research pass over each vendor’s own documentation; hedges are printed in the cards wherever a vendor hasn’t documented a feature. Nothing on this page calls a model or sends your text anywhere.