學術研究神器 · AI 論文發想框架 · Spark2Groundwork

The Researcher's Power Tool · An AI Framework for Paper Ideation

把一個發想，變成一份站得住的研究提案。 Turn a spark of an idea into a research proposal that holds up.

English edition ｜ 繁體中文版 ｜ What changed in each release

這份說明以英文撰寫。中文使用者請點上面的「繁體中文版」，那裡有完整的中文說明。

---

What this is

A set of files you copy into your project folder. They give your AI assistant a way of working, and they give you a way of checking what it did.

It starts from one assumption: the AI will be wrong, and it will be wrong in the places where it sounds most certain. Every check in here exists because that actually happened to someone.

The questions it helps you answer
Which page does this sentence rest on?
What would it take to prove this idea wrong?
What did the AI just change, and have I looked at it?

⛔ It does not promise your research is correct. It promises something narrower and more useful: when something goes wrong, it leaves a trace instead of quietly becoming your conclusion.

---

How it works

Four ideas hold the whole thing together.

1. Getting from a source to a sentence takes several steps, and they break differently. One of those steps — is this passage actually in the paper? — can be settled by comparing text, at no cost. That one is automated, and the rest of the design leans on it. Two other steps — does the passage really support the claim? and can you generalise from that study to yours? — ⛔ are deliberately left to you. A check that fires on correct work teaches you to ignore it, which is worse than having no check at all.

2. Two records, two different jobs. One tracks your ideas and what would prove each of them wrong. The other tracks your sentences and the exact passage each one rests on.

3. Mistakes come in families. Writing down one mistake does not help, because the next one looks different. What you write down is the pattern — so you recognise it next time.

4. You are the one who decides. This is not caution; it is what happened. Across two earlier projects, the user caught more genuinely new errors than any other part of the system, and the most effective thing they did was bring in information the AI did not have.

---

There is no perfect framework — only one that gets better as you use it

🔴 This is the idea the whole thing is built on.

No framework arrives already suited to your topic, your AI and your working habits. ⛔ We are not going to pretend otherwise.

It is built on a different premise: every hole you fall into becomes a railing for next time.

What you do	What the framework grows
Catch the AI getting something wrong and say "log that"	One more entry in your project's incident log
After a few entries, ask the AI to find the repeating patterns	You see a shape that will happen again
You decide whether it becomes a rule	One more working rule — specific to your project

⚠️ The middle step is not optional. One mistake on its own does not help, because the next one will not look the same; what is worth having is the shape that repeats.

⛔ The last step is always your decision, never the AI's — because that rule will bind both of you afterwards.

After a while your copy will not look like anyone else's. 🔴 That is not drift. That is the point.

⚠️ It is also why upgrading is safe, for a reason that is easy to read backwards: the upgrade tool recognises nine names of its own — five framework folders and four framework files — and ⛔ refuses everything else. Your ledgers, your papers, your incident log, PROJECT.md, and any folder you created yourself are all left alone. You never have to register them anywhere.

---

Getting started

You do not need to know Git, and there is nothing to install for this step.

1. Click the green Code button at the top of this page, then Download ZIP.
2. Unzip the file you downloaded.
3. Inside it you will find two folders. Take the one in your language:
   • Spark2Groundwork_en — English
   • Spark2Groundwork_zh — 繁體中文
4. Copy that folder to wherever you keep your work, and rename it to your project name. For example bilingual-memory-study.
5. ⛔ Delete the rest of the download. You only need the one folder.
6. Open SETUP.md inside your new folder and follow it.

What you will need later: Python (free, from python.org), Git (free, from git-scm.com), and whichever AI you already use. SETUP.md walks you through installing them.

---

Pointing your AI at the folder

The framework is just files, so any AI that can read your folder can use it. Three common setups, with the steps as of this writing:

Claude — use Cowork

1. Open the Claude desktop app.
2. In the left panel, find Projects and click +.
3. Choose Use an existing folder on your computer.
4. Pick your project folder, give the project a name, and click Create.

Claude can then read and write the files directly, and can run the checks for you.

Gemini — use Antigravity

1. Open Antigravity.
2. Click Select Project → New Project.
3. Use Add Folder to add your project folder, then create the project.
4. Chat with the agent in the main panel. Open IDE gives you a full editor if you want one.

A project can hold more than one folder, so you can add related material alongside it.

ChatGPT — use Work mode in the desktop app

1. Open the ChatGPT desktop app and switch to the Work tab at the top.
2. Click Select project below the input box, then New project.
3. In the dialog that opens, type a project name.
4. Under Source folder, click the box offering to add a folder ChatGPT can read and edit. A file picker opens, titled Select Project Root — choose your project folder.
5. Click Create project.

Work can then read and write files in that folder, and the link stays put between sessions — ⛔ you do not re-upload anything.

⚠️ Local file access depends on your plan and, in an organisation, on your workspace settings. If the option to link a folder is not offered, that is why.

⚠️ Web ChatGPT is a different thing. There, a project holds uploaded copies rather than a live folder, with a file limit (5 on Free, 25 on Go/Plus, 40 on the higher plans). It still works, with more copying and pasting — profiles/PROFILE_chat_only.md in your folder is written for exactly that case.

Already using a coding agent?

Claude Code and ChatGPT's Codex both work with this framework too, and they are the most direct fit — they live in a folder, read files and run commands natively.

⚠️ We have not tested either of them with this framework, so we are not giving you steps. Point them at the folder the way you normally would, and start from SETUP.md.

⛔ One thing to know before you do: a coding agent will happily edit files and run commands without pausing. The protection in this framework comes from the checkpoint habit — press snapshot before you hand out work, and again after you have reviewed it. Nothing enforces that for you.

⚠️ Menus change. If what you see does not match the steps above, look for the wording that means the same thing — "add a folder", "link a folder", "project instructions".

---

What is in the box

Spark2Groundwork_en/     the whole framework, in English
Spark2Groundwork_zh/     完整框架，繁體中文

The two are independent and say the same things. Take one; you do not need both.

Inside either one:

🔴 Everything splits into two piles, and the line is who owns a file.

Yours — an upgrade never touches any of it:

PROJECT.md	🔴 The only file you have to write yourself
my/MY_RULES.md	Your working rules. The framework's are copied into §1; the P-xx you write go in §2
my/MY_INCIDENTS.md	What actually went wrong in this project
my/MY_INDEX.md	An index of everything here that is yours — generated, never hand-written
my/tools/	Scripts you wrote. ⛔ Not in scripts/, which is replaced on upgrade
corpus/	Your source PDFs go here, along with your bibliography
corpus_md/	The plain text pulled out of those PDFs, for checking quotations
ledgers/	Where your ideas and your quoted evidence are recorded
handoffs/	Handoff packets: what each round did
NEXT_SESSION_MEMO.md	Working state, overwritten each round
governance_config.json	You decide where the AI may write

The framework's — if it breaks, download it again:

SETUP.md	The setup guide. Start here
INITIALIZE_PROMPT.md	The text you paste to your AI the first time
governance/	The rules the AI works under, and the two ledgers' field specifications
policy/	Rules for particular topics: sources, handovers, model identity, outside tools
profiles/	Pick the one that matches how you work
prompts/	Ready-made instructions you can paste
scripts/harness/	The automatic checks
docs/	The figures
file_index.md	The framework's own index. ⛔ Do not register your documents here
Three buttons	snapshot / review changes / check update (.bat for Windows, .command for Mac)

🔴 Why the split: a file may have exactly one owner. Mix the framework's content with yours and an upgrade will lose one of them. ⚠️ So your rules live in my/, ⛔ not in governance/ — and when the framework adds a rule, a tool copies the new text into your file verbatim while leaving the rules you wrote alone.

---

Versions

Each release is tagged in Git, and the folder names carry no version number — if they did, every path and bookmark would break on each release.

See CHANGELOG.md for what changed, and Releases for the downloads.

To find out whether you are up to date, press the check update button in your project folder.

---

Where this came from

Two real research projects, and forty-odd recorded mistakes between them. Every check and every rule here carries the case that caused it.

⚠️ Some of the mistake patterns listed in governance/Incident_Log.md are marked [inherited], [framework's own] or [predicted]. Those have not happened in your project. They are listed so you know what to watch for — ⛔ not so you can assume you are already safe from them.

---

Licence

MIT — see LICENSE. The framework files are covered; the research you produce with them is yours.