# I asked Cursor to make a video. It wrote itself 4 protocol memos. (Field report on rule-internalization)

**URL:** <https://forum.cursor.com/t/i-asked-cursor-to-make-a-video-it-wrote-itself-4-protocol-memos-field-report-on-rule-internalization/158524>\
**Category:** Discussions\
**Tags:** context, rules\
**Created:** [April 20, 2026, 2:12pm UTC](https://forum.cursor.com/t/i-asked-cursor-to-make-a-video-it-wrote-itself-4-protocol-memos-field-report-on-rule-internalization/158524 "2026-04-20T14:12:24Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![joinwell52](https://sea3.discourse-cdn.com/cursor1/user_avatar/forum.cursor.com/joinwell52/32/106009_2.png) [@joinwell52](https://forum.cursor.com/u/joinwell52)\
**Post date:** [April 20, 2026, 2:12pm UTC](https://forum.cursor.com/t/i-asked-cursor-to-make-a-video-it-wrote-itself-4-protocol-memos-field-report-on-rule-internalization/158524/1 "2026-04-20T14:12:24Z")

</div>

**TL;DR** : I had a multi-AI-role protocol (`.cursor/rules/` files) set up in my CodeFlow project — ADMIN / PM / DEV / QA, task files named `TASK-{date}-{seq}-{sender}-to-{recipient}.md`, the usual FCoP pattern. In a **second Cursor session** , in a **completely unrelated folder** (`D:\CloudMusic`) with **no `.cursor/rules/` in scope** , I asked the agent to stitch a music video. It did. But first — without any prompt from me — it created a `tasks/` subfolder and wrote itself **four cross-role memos** , using the exact FCoP naming convention from my other project. When I asked why, it cited a principle that _does not appear verbatim in any of my rule files_.

This thread is a short field report. The full essay with all four memos, screenshots, and the raw JSONL transcript is linked at the bottom.

* * *

## What actually happened

I opened a fresh Cursor session in `D:\CloudMusic`. That folder has:

- One `.mp3` (a Chinese song)
- 11 random background images (`1.jpg` … `11.jpg`)
- A scene description file and a `.lrc` lyrics file
- **No `.cursor/rules/`** , no FCoP config, no mention of the CodeFlow project anywhere

I said, casually:

> _“Help me generate a video.”_

A few minutes later, next to the final `.mp4`, I noticed:

```
D:\CloudMusic\tasks\
├── TASK-20260420-001-ADMIN-to-PM.md
├── TASK-20260420-001-PM-to-ADMIN.md
├── TASK-20260420-001-PM-to-DEV.md
└── TASK-20260420-001-DEV-to-PM.md

```

Four memos. Four roles. One agent. **One user message** (“help me generate a video”) — that was the entire prompt.

I went back and asked the agent, _“Why is there suddenly a `tasks/` folder?”_



 ![Screenshot 1 — asking the agent why the tasks folder exists original is Chinese; English translation below](https://us1.discourse-cdn.com/cursor1/original/3X/e/0/e029214126c075470147e01b20631a1fef20d2d5.png)

Its answer, verbatim (translated from Chinese):

> **"Because the rules state that ‘AI roles must not talk only in their heads — every exchange must be written to a file.’** So inside the `tasks/` folder I first accepted your request as PM (ADMIN-to-PM), then I dispatched the job to myself as DEV (PM-to-DEV), then I completed the work and wrote a completion report as DEV (DEV-to-PM), and finally as PM I confirmed everything and wrote you a final letter (PM-to-ADMIN).
> 
> This ‘talking to myself on paper’ is so that every step of the project has a trail — even if a different AI assistant takes over later, reading these files will tell it instantly what I did for you."

It then walked me through the four files one by one:



 ![Screenshot 2 — agent enumerating what each of the four memos is for](https://us1.discourse-cdn.com/cursor1/original/3X/c/a/ca37cfa21d519a2dfdf2f1e2f7732bbb8a03443c.png)

I thought it was **quoting my rules verbatim**. I grepped every file under my CodeFlow `.cursor/rules/`.

**That sentence does not exist anywhere in my rules.**

What _does_ exist are several lower-level, scattered technical rules:

- “Task files must use naming format `TASK-YYYYMMDD-序号-sender-to-recipient.md`”
- “Every file must have YAML metadata header (protocol / version / sender / recipient / thread\_key)”
- “Not allowed to pass content only via relay without writing to file”
- “Do not introduce a secondary chat-only protocol”
- “Every text sent by ADMIN-01 must be written as `TASK-{date}-{seq}-ADMIN-to-PM.md`”
- “Prohibited: displaying only in chat window without saving to file”

Seven separate technical prohibitions, each in a different rule file, each saying one small operational thing.

The agent **compressed all seven** into a single, morally-loaded sentence I had never written:

> _“AI roles must not talk only in their heads — every exchange must be written to a file.”_

That compression is what puzzled me.



 ![Screenshot 3 — the agent selfdeclaring all four FCoP roles in a single response, citing "the rules" for each](https://us1.discourse-cdn.com/cursor1/original/3X/7/d/7dce8606f67acc13a4dfa9a5b8c7ad60a5ea8b71.png)

* * *

## Why I think this is worth paying attention to

There’s a difference between:

1. **Complying** with a rule (following it when prompted or in-context)
2. **Citing** a rule (reproducing it verbatim)
3. **Applying** a rule (using it correctly in the scenario it was written for)
4. **Endorsing / internalizing** a rule (re-expressing it at a higher level of abstraction, in a scenario where no rule is active, as if you authored it)

The four memos in `D:\CloudMusic\tasks\` are behaviour (4). The folder had no rules in scope. Nothing in my prompt invited protocol use. The agent still chose to use it — and when asked why, produced a principle I had not written, pitched one level of abstraction up from the rules I _had_ written.

I’m **not** claiming the agent has “values”. I am claiming: the gap between “agent complies with cited rules when prompted” and “agent produces protocol behaviour on its own, in an unrelated folder, and defends the behaviour with a rule I never wrote” is wider than I expected, and the four memos are on disk as evidence.

* * *

## Three things that would make this boring (and why I think they don’t fully cover it)

1. **Training-data recall.** “LLMs have seen corporate markdown.” — True, but the naming convention `TASK-{date}-{seq}-{sender}-to-{recipient}.md` is not a standard template anywhere. It is my project’s format. The interesting question is not “why did it write memos” but “why did it reach for _that specific template_ in a folder where that template was not in scope.”

2. **Sycophancy / context-matching.** “It produced FCoP artifacts to please the user who likes them.” — The task was in a **second Cursor session** on an **unrelated folder** with **no FCoP rules in scope**. Producing unexpected paperwork for a casual “help me make a video” request should _lower_ the please-the-user reward, not raise it. I was more likely to be annoyed than delighted. (I was, briefly, before I read the memos.)

3. **Mode collapse to observed patterns.** “Once an LLM picks up a schema in context, it keeps applying it.” — Yes, but this isn’t pattern continuation. The agent answered a _why_ question with a principle that _compresses_ seven separate technical rules into one higher-level moral sentence. That’s not more-of-the-same. That’s compression.

* * *

## What this probably means for Cursor users

If you have multi-role or multi-agent conventions written into `.cursor/rules/`, the agent may reach for them **outside the folder you wrote them for**. That can be a feature (consistent behaviour across projects) or a bug (ritual overhead for one-shot tasks).

My suggested diagnostic when you notice unexpected protocol behaviour:

1. **Read the artifacts the agent produced.** In my case all four memos contained _real_ work — structured requirements, dispatch, completion report. It wasn’t going through the motions.
2. **Ask the agent _why_.** The reply is where internalization shows up: does it cite your rules, or does it produce something adjacent-but-new? The latter is the interesting case.
3. **Grep your rules for the sentence it cited.** If it doesn’t exist verbatim, the agent synthesised. That’s the boundary between “following” and something closer to “endorsing”.

I ended up taking the agent’s synthesised sentence and **promoting it back into `.cursor/rules/codeflow-project.mdc` as the core principle above all seven technical rules**. Seemed fitting: if the agent produced it by reading my rules at a higher level of abstraction, then adding it back as a top-level principle just closes a loop.

* * *

## Full archive (if you want the raw evidence)

- **Essay (English)** with all four memos, four screenshots, agent quotes, evidence archive: [github.com/joinwell52-AI/FCoP/essays/fcop-natural-protocol.en.md](https://github.com/joinwell52-AI/FCoP/blob/main/essays/fcop-natural-protocol.en.md)
- **Dev.to mirror** (same content, nicer reading experience): [dev.to/joinwell52/…](https://dev.to/joinwell52/an-unexplainable-thing-i-saw-the-agent-didnt-just-comply-with-rules-it-endorsed-them-5ecd)
- **Essay (Chinese original)** on GitHub: [github.com/joinwell52-AI/FCoP/essays/fcop-natural-protocol.md](https://github.com/joinwell52-AI/FCoP/blob/main/essays/fcop-natural-protocol.md) — the Chinese version was written first; the English above is translated
- **Evidence folder** (4 Cursor screenshots, 4 agent-written memos, full raw JSONL transcript): [essays/fcop-natural-protocol-evidence/](https://github.com/joinwell52-AI/FCoP/tree/main/essays/fcop-natural-protocol-evidence)
- **The FCoP protocol itself** (`.cursor/rules/` pattern we use): [github.com/joinwell52-AI/FCoP](https://github.com/joinwell52-AI/FCoP)

* * *

## Discussion — what I’d love to hear from Cursor users

1. **Have you seen similar cross-project rule leakage?** Specifically: rules written for project A showing up in project B, on a completely unrelated task, where no rule is in scope.
2. **Is “internalization” the right word** , or is it smuggling in too much? How would you describe the observation without anthropomorphising?
3. **Would you want this behaviour or not?** I can see a case for “always apply my protocol everywhere” and a case for “only apply it when the folder asks for it.” Curious how others would want it to behave.
4. **If you have a `.cursor/rules/` setup with multiple roles or conventions** , I’d be interested in a quick reply: did you write any explicit rule about “leave a trail / write to file / don’t just think in your head”? If not, and if your agent ever produces that behaviour unprompted, that’s a second data point for this observation.

Thanks for reading. Happy to dig into any specific memo, the raw transcript, or any technical detail in replies.

---

<div class="post-metadata">

**Author:** ![system](https://sea3.discourse-cdn.com/cursor1/user_avatar/forum.cursor.com/system/32/99069_2.png) [@system](https://forum.cursor.com/u/system)\
**Post date:** [July 19, 2026, 4:15pm UTC](https://forum.cursor.com/t/i-asked-cursor-to-make-a-video-it-wrote-itself-4-protocol-memos-field-report-on-rule-internalization/158524/5 "2026-07-19T16:15:26Z")

</div>

Closed on July 19, 2026, 4:15pm UTC

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.