---
title: "Build ChatGPT Custom Instructions From Your Chat History With Jev and Claude"
date: "2026-09-25"
video_id: "7689525409691553037"
tiktok_url: "https://www.tiktok.com/@angusthenontechnical/video/7689525409691553037"
creator: "@angusthenontechnical"
category: "Marketing"
tags: ["custom instructions", "chatgpt export", "jev", "claude", "mcp"]
excerpt: "How to export your ChatGPT history, grade each chat with TypeSafe's Jev, and have Claude write custom instructions. What it costs and what the video gets wrong."
---
A TikTok from Angus the Nontechnical (@angusthenontechnical), captioned "Jarvis make my content go mega viral," shows a workflow for writing better AI custom instructions from your own chat history. You export every ChatGPT conversation, have TypeSafe's Jev model flag the chats where the AI went wrong, and have Claude read only those to write a custom instructions block. The workflow is real and cheap to run. It has nothing to do with virality, and it was not "automated by ChatGPT."

## How the custom instructions workflow works

The pipeline has three steps:

1. **Export your history.** ChatGPT and Claude both let you download all your conversations as a file, `conversations.json` in ChatGPT's case.
2. **Grade each chat with Jev.** Jev asks one yes-or-no question per chat, such as "too verbose?", "got it wrong?" or "made me repeat myself?" In the video, the output summary for 10 chats reads `"got_it_wrong": 3, "got_it_right": 7`.
3. **Have Claude read only the flagged chats.** Claude writes the instructions from those.

The prompt shown on screen begins "This is my whole ChatGPT history. Reverse engineer the custom instructions that turn you into my Jarvis," and includes "Keep patterns that show up in 3+ chats," "Paste my own words in as a voice sample," and "Under 1,500 characters. A ChatGPT and Claude version."

The result is a 1,433-character block opening "You are my chief of staff," built from patterns with counts such as "Pick one, don't list (11 chats)," "Human voice, not corporate or AI (10 chats)" and "Lead with the point, no preamble (5 chats)." He pastes it into the custom instructions panel.

The design keeps costs down: Jev handles the cheap per-chat questions, and the expensive model reads only a subset.

## How to export your ChatGPT and Claude history

[OpenAI's help article](https://help.openai.com/en/articles/7260999-how-do-i-export-my-chatgpt-history-and-data) gives the path: profile menu, Settings, Data controls, Export data, Export. Claude's export is under Settings, Privacy, Export data.

The video skips the wait. OpenAI emails the archive, processing is reported to take up to seven days, and the download link expires 24 hours after delivery. Claude's export also arrives by email with a 24-hour link that must be opened while signed in. Request the export first.

## What is Jev, and is it a small model?

The video says "You gotta attach any small model to it, like the Jev model." Jev is not a small chat model. [TypeSafe](https://typesafe.ai) describes it as its "first public System One Model, optimized for automation," returning "typed decisions with calibrated probabilities." The [documentation](https://docs.typesafe.ai/) lists three question types: Choice and Score return a value with probabilities and a confidence number, and Noul returns a value between 0 and 1. It does not write prose, so a small LLM is not a drop-in replacement.

The video connects it through [`jev-mcp-server`](https://pypi.org/project/jev-mcp-server/), installed with `uvx jev-mcp-server install claude-code` after getting a key from the TypeSafe console. The [models page](https://docs.typesafe.ai/models) confirms `jev-latest` is an alias for `jev-1.13.0`, the model named in the video's output.

## What does Jev cost?

TypeSafe lists input at $42 per billion tokens ($0.042 per million) with free output. Rate limits on `jev-1.13.0` are 100,000 tokens per second and 80 requests per second.

TypeSafe claims a "238x lower input price than Claude Fable 5.1." That checks out: [Anthropic lists](https://platform.claude.com/docs/en/docs/about-claude/models/overview) Claude Fable 5.1 at $10 per million input tokens, and 10 divided by 0.042 is 238.1. The current Claude lineup is Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5. The video does not say which Claude model it used.

## Did ChatGPT automate personal AI assistants?

No. The video opens "These AI personal Jarvis's just got automated by ChatGPT." OpenAI's only part is the data export, which has existed for years. TypeSafe's Jev does the grading and Claude does the writing.

## Does this make content go viral?

No. The workflow uses no social platform API, engagement data or posting tool. Its only input is the user's own chat history. The video's first title card reads "There was never a Jarvis," and its closing line is "most people don't need to spend the time building this."

## Where is the "Please don't buy a Mac Mini" quote from?

The video shows a real article from The Next Web, ["Claude AI agents are driving record Mac mini demand,"](https://thenextweb.com/news/claude-mac-mini-openclaw-ai-agents-shortage) published May 11, 2026 by Alina Maria Stan, about agents built on OpenClaw.

The next frame circles a quote that reads as part of the same article: "Please don't buy a Mac Mini... You can deploy this on Amazon's Free Tier." That article does not contain it. The quote is a [post on X](https://x.com/steipete/status/2015252888528859603) by Peter Steinberger, the Austrian programmer who created [OpenClaw](https://en.wikipedia.org/wiki/OpenClaw), first released as Warelay in November 2025.

## What is the character limit for ChatGPT custom instructions?

The prompt targets under 1,500 characters, which is ChatGPT's long-standing custom instructions limit and still the documented limit on Free and Go accounts.

## Bottom line

Exporting your chat history and grading it with a cheap classifier before an LLM writes your instructions is a sensible way to get custom instructions built on real patterns. It costs fractions of a cent per chat in Jev fees. It will not make anything go viral, and OpenAI's contribution is the export button.

## Sources

- [The Next Web on Mac mini demand](https://thenextweb.com/news/claude-mac-mini-openclaw-ai-agents-shortage) - the article shown in the video
- [Peter Steinberger on X](https://x.com/steipete/status/2015252888528859603) - source of the Mac Mini quote
- [OpenClaw on Wikipedia](https://en.wikipedia.org/wiki/OpenClaw) - creator and release history
- [TypeSafe AI](https://typesafe.ai) - Jev description and pricing
- [TypeSafe models documentation](https://docs.typesafe.ai/models) - model IDs and rate limits
- [TypeSafe documentation index](https://docs.typesafe.ai/) - Choice, Score and Noul
- [jev-mcp-server on PyPI](https://pypi.org/project/jev-mcp-server/) - the MCP package
- [Anthropic models overview](https://platform.claude.com/docs/en/docs/about-claude/models/overview) - Claude lineup and pricing
- [OpenAI: exporting your ChatGPT history](https://help.openai.com/en/articles/7260999-how-do-i-export-my-chatgpt-history-and-data) - export steps


## Video description and verification notes

**The artifact is a 62.135-second (1:02) vertical clip at 1080x1920, HEVC Main profile in yuv420p at a constant 30 fps across 1,864 video frames, with an HE-AACv2 audio track at 44.1 kHz stereo and 48,158 bps, totalling 6,865,124 bytes (6.5 MiB) in an MP4 container at an overall 883,897 bps. It was posted 2026-09-25 at 17:26:29 UTC by `angusthenontechnical`, channel nickname "Angus the Nontechnical". At capture on 2026-10-06 it showed 25,600 views, 1,352 likes, 12 comments and 283 reposts. The audio is listed as "original sound" credited to Angus the Nontechnical, meaning his own lavalier voiceover with no licensed music bed. I read 21 of the 31 extracted frames against a 258-word transcript. The opening two frames carry a white title card reading "There was never a Jarvis" over a dark force-directed graph UI showing nodes labelled "Finances", "Clients", "Marketing/Growth", "TECH", "Sales", "Communications", "Payments Pulse", "Onboard new clients", "Watch processor he...", plus vendor nodes "Attio", "Notion", "Slack", "Zeroentropy", "Supabase", and a top-edge input hint "...or drop documents". Frames 3 and 4 show a The Next Web article page: logo "TNW", nav items "Latest / Deep tech / Sustainability / Ecosystems", a black bar reading "This article was published on May 11, 2026", kicker "ANTHROPIC", headline "Claude AI agents are driving record Mac mini demand", and byline time "May 11, 2026 - 2:59 pm", with a hand-drawn red circle around "...sold out across the United States" and a second red circle in frame 4 around the quote "Please don't buy a Mac Mini," he wrote. "You can deploy this on Amazon's Free Tier." Frame 7 shows the ChatGPT account menu for user "angussewell" marked "Pro", with "Personalization / Profile / Settings / Help / Log out" and a red circle on "Settings", then a second panel numbered 2 showing the settings sidebar ("Plugins, Voice, Billing, Usage, Analy..., Data c..., Cloud..., Storag..., Safety, Security a..., Parental c..., Trusted co..., Ac..., Ke...") with rows "Improve the model for everyone  Off", "Information shared with apps", and buttons "Reset", "Manage", "Manage", "Archive all", "Delete all", "Export", the last circled in red. Frame 8 repeats the exercise in Claude: a menu with "Settings / Language / Get help / View all plans / Get apps and extensions / Claude Academy (New) / Learn more / Log out", then a Settings sidebar reading "General, Account, Privacy, Billing, Usage, Capabilities, Memory, Claude Code, Network, Claude in Chrome" with "Privacy" circled and an "Export data" button circled at bottom right. Frames 9 and 10 show a chat composer holding an attachment chip labelled "conversations.json  JSON" above an eight-point prompt beginning "This is my whole ChatGPT history. Reverse engineer the custom instructions that turn you into my Jarvis," whose numbered items include "Keep patterns that show up in 3+ chats", "Write it as latent space engineering, not rules", "Paste my own words in as a voice sample", and "Under 1,500 characters. A ChatGPT and Claude version." Frames 11 and 12 show a documentation page with an `h2` "Quickstart" and a code block reading `# 1. Get a key: https://console.typesafe.ai/sett...`, `# 2. Install into your client`, `uvx jev-mcp-server install claude-code`, plus install fallbacks `curl -LSsf https://astral.sh/uv/install.sh | sh`, `brew install uv`, `pip install uv`, `pip install jev-mcp-server`, and a "Tools" table with a "choice" row. Frames 13 to 17 show a flow diagram: `conversations.json` captioned "every chat you ever had", a divider "ONE QUESTION PER CHAT", three "Jev" rows asking "too verbose? yes or no", "got it wrong? yes or no", "made me repeat myself? yes or no", and a "Claude" row reading "reads only the flagged chats", under a black box reading "custom instructions / built on real counts". Frame 20 is raw JSON: `{"tool": "jev_classify", "model": "jev-latest", "results": [...]}` with per-chat objects such as `{"id": "chat_0427", "classification": "got_it_wrong", "top_probability": 1}` and a closing `"summary": {"items": 10, "auto": 10, "by_class": {"got_it_wrong": 3, "got_it_right": 7}}`. Frame 21 renders the same data as a table with columns "CHAT / TOO VERBOSE / GOT IT WRONG / CLAUDE" and rows chat 0412 through chat 0467 resolving to "read" or "skip". Frames 18, 19 and 22 show the output: a header "Claude version (1,433 chars)" over a prompt beginning "You are my chief of staff: a direct-response copywriter with Harry Dry's eye for the one sharp reframe, who debugs like a staff engineer", a "Patterns that showed up in 3+ chats" list with counts ("Pick one, don't list (11 chats)", "Human voice, not corporate or AI (10 chats)", "Verify before it goes public (5 chats)", "Commit to a take, be blunt (5 chats)", "Stay in scope and respect his draft (3 chats)", "Lead with the point, no preamble (5 chats)"), and finally that text pasted into a "Custom instructions" settings panel with a style selector reading "Pet / Default" and Cancel and Save buttons. The last nine frames are talking-head shots by a window with no further UI.**

## The title and the video describe two different things

The post is captioned "Jarvis make my content go mega viral  #ai #buildwithai #vibecoding". Nothing in the 62 seconds touches content distribution, posting schedules, hook testing, or view prediction. The first visible card contradicts the caption directly: it reads "There was never a Jarvis". The spoken open is "These AI personal Jarvis's just got automated by ChatGPT," and the close is "Setting up your custom instructions like this? Great first start."

There is also no social platform API anywhere in the build. No TikTok Display API, no TikTok Research API, no Instagram Graph API, no YouTube Data API. The only data source is the operator's own chat history. No part of this pipeline reads, predicts, or influences platform performance, so there is no virality claim to assess. The caption is hook-writing detached from the build.

## "Automated by ChatGPT" is the wrong attribution

The transcript asserts that these agents "just got automated by ChatGPT." The frames show three separate vendors doing three separate jobs, and OpenAI performs only the first.

OpenAI supplies the export. Per [OpenAI's own help article](https://help.openai.com/en/articles/7260999-how-do-i-export-my-chatgpt-history-and-data), the path is profile menu, Settings, Data controls, Export data, Export, which matches the circled UI in frame 7. TypeSafe supplies the per-chat grading through the Jev API, called over MCP. Anthropic's Claude reads the flagged subset and writes the final instruction block, which is then pasted back into a Claude custom instructions panel in frame 22.

OpenAI shipped no feature here. A data-portability export that has existed for years is being used as a corpus. The operator built a three-vendor pipeline on top of that export.

## Jev is a typed-decision model, not "a small model" in the usual sense

The transcript says "You gotta attach any small model to it, like the Jev model." The on-screen artifact is `jev-mcp-server`, an MCP server that fronts TypeSafe's hosted API, and the JSON in frame 20 names `"model": "jev-latest"`.

That model ID is real. [TypeSafe's models documentation](https://docs.typesafe.ai/models) lists `jev-1.13.0` as the current version and confirms `jev-latest` as a valid alias pointing at it. [TypeSafe's own site](https://typesafe.ai) describes Jev as its "first public System One Model, optimized for automation," returning "typed decisions with calibrated probabilities." [The documentation index](https://docs.typesafe.ai/) names exactly three primitives: Choice, Score, and Noul. Choice and Score return a value plus probabilities plus a confidence number. Noul returns a value between 0 and 1.

Calling it "a small model" misleads. Jev does not generate prose and cannot be swapped for a small LLM. It answers an enumerated question and returns a distribution. The flow diagram in frames 13 to 17 is a faithful picture of that constraint: each Jev row is a single yes-or-no question posed once per chat, and the LLM step sits downstream.

Pricing and limits are published. TypeSafe lists input at $42 per billion tokens, which is $0.042 per million tokens, with output tokens free. Rate limits on `jev-1.13.0` are documented at 100,000 tokens per second and 80 requests per second, returning `429 Too Many Requests` above that. TypeSafe's marketing claim of "238x lower input price than Claude Fable 5.1" checks out arithmetically against [Anthropic's published pricing](https://platform.claude.com/docs/en/docs/about-claude/models/overview) of $10 per million input tokens for Claude Fable 5.1, since 10 divided by 0.042 is 238.1. The companion "193.6x Faster" claim has no methodology attached on the page and I did not verify it.

For the Anthropic half of the pipeline: the current lineup is Claude Fable 5.1 (`claude-fable-5-1`), Claude Opus 5.5 (`claude-opus-5-5`, the recommended starting point), Claude Sonnet 5.5 (`claude-sonnet-5-5`) and Claude Haiku 4.5 (`claude-haiku-4-5`). Claude Opus 5 and Claude Sonnet 5 remain available as legacy. Every current model takes text and image input and returns text only. The video names no specific Claude model, so nothing there needs correcting.

## The Mac mini article is real, but the circled quote is not from it

Frame 3 shows a genuine page. The Next Web published "Claude AI agents are driving record Mac mini demand" on May 11, 2026 at 2:59 pm, written by Alina Maria Stan. The article's subject is Tyler Cadwell of Everything Etched, who built an agent called Etchie on OpenClaw, and it cites roughly 247,000 GitHub stars for the project plus Tim Cook's Q2 2026 earnings-call remarks attributing Mac inventory tightness to supply rather than demand.

Frame 4 is a different document. I fetched the TNW article and it contains no mention of Steinberger, no ElevenLabs or OpenTable anecdote, and no line reading "AGI is here and 99% of people have no clue." The quote circled in red does exist, but as a post by Peter Steinberger on X: "Please don't buy a Mac Mini, rather sponsor one of the many contributors of @openclaw You can deploy this on Amazon's Free Tier." Steinberger is the Austrian programmer who created OpenClaw, originally released as Warelay in November 2025. Two sources have been edited together into what reads as one continuous scroll, so a viewer would reasonably attribute the quote to TNW. That attribution is wrong.

## The 1,500-character ceiling in the prompt is a floor on some plans and may be stale on others

Item 8 of the on-screen prompt reads "Under 1,500 characters. A ChatGPT and Claude version," and the delivered Claude output is headed "Claude version (1,433 chars)."

1,500 is the long-standing ChatGPT custom instructions field limit and remains the documented limit on Free and Go accounts. Several secondary write-ups from mid-2026 report that paid plans now accept around 5,000 characters. I could not confirm that against OpenAI's own help centre, which returned HTTP 403 to automated fetches, and the secondary reports cite no OpenAI announcement, so treat the expanded number as unverified. Frame 7 shows the operator's ChatGPT account labelled "Pro", so if a higher paid-plan limit exists he is constraining himself below it for no stated reason.

## The export step has two timing facts the video skips

The video presents the export as instant: "Export all of your AI data," then straight to a chat composer with `conversations.json` attached. Two documented delays sit between those shots. OpenAI delivers the archive by email and the download link expires 24 hours after delivery, with processing reported to take up to seven days. Claude's export follows the same pattern, delivered by email from Settings, Privacy, Export data, with a link that also expires after 24 hours and must be opened while signed into the requesting account.

Anyone copying this workflow should request the export first and do everything else while waiting.

## Key Takeaways

- **Verified:** The pipeline on screen is real. `jev-mcp-server` exists, installs via `uvx jev-mcp-server install claude-code`, and `jev-latest` is a documented alias for `jev-1.13.0`. The export paths circled in frames 7 and 8 match both vendors' documented settings routes.
- **Verified:** TypeSafe publishes Jev input pricing at $42 per billion tokens with free output, and rate limits of 100,000 tokens per second and 80 requests per second. Its "238x lower input price than Claude Fable 5.1" claim is arithmetically correct against Anthropic's $10 per million input token price for Fable 5.1.
- **Correction:** The title promises content virality. The video contains no virality mechanism, no engagement prediction, and no social platform API of any kind. Its own first title card reads "There was never a Jarvis."
- **Correction:** "These AI personal Jarvis's just got automated by ChatGPT" misattributes the work. OpenAI contributes only a long-standing data export. The grading is TypeSafe's Jev and the writing is Claude.
- **Correction:** The red-circled "Please don't buy a Mac Mini" quote is presented as part of the TNW article shown one frame earlier. It is not in that article. It is a post by Peter Steinberger, creator of OpenClaw, on X.
- **Partial correction:** Calling Jev "a small model" implies an interchangeable small LLM. TypeSafe describes it as a System One model returning typed decisions with calibrated probabilities across three primitives, Choice, Score and Noul. It does not produce prose and is not a drop-in for a small chat model.
- **Unverified:** Reports that ChatGPT's custom instructions limit rose from 1,500 to around 5,000 characters on paid plans. OpenAI's help centre blocked automated fetches and the secondary sources cite no official announcement. The 1,500 figure remains correct for Free and Go.
- **Unverified:** TypeSafe's "193.6x Faster" claim. No benchmark methodology is published alongside it.
- **Context:** The design is cost control. Jev answers one yes-or-no question per chat at fractions of a cent, and Claude reads only the flagged subset. Running a frontier model over an entire `conversations.json` is the obvious alternative and the expensive one.
- **Context:** The creator's own closing line is the honest summary of the video: "most people don't need to spend the time building this."

## Resources

- [Claude AI agents are driving record Mac mini demand, The Next Web](https://thenextweb.com/news/claude-mac-mini-openclaw-ai-agents-shortage). Establishes the real publication date, author and contents of the article shown in frame 3, and confirms it does not contain the quote circled in frame 4.
- [Peter Steinberger on X](https://x.com/steipete/status/2015252888528859603). The actual source of "Please don't buy a Mac Mini... You can deploy this on Amazon's Free Tier."
- [OpenClaw, Wikipedia](https://en.wikipedia.org/wiki/OpenClaw). Establishes Steinberger as the creator, the November 2025 release under the name Warelay, and the 247,000 GitHub stars figure.
- [TypeSafe AI](https://typesafe.ai). Describes Jev as a System One Model returning typed decisions with calibrated probabilities, and carries the $42 per billion input tokens price and the 238x comparison.
- [TypeSafe models documentation](https://docs.typesafe.ai/models). Confirms `jev-1.13.0`, the `jev-latest` alias, free output tokens, and the rate limits.
- [TypeSafe documentation index](https://docs.typesafe.ai/). Confirms the three question types, Choice, Score and Noul, and what each returns.
- [jev-mcp-server on PyPI](https://pypi.org/project/jev-mcp-server/). Confirms the package shown in the Quickstart frames and the `console.typesafe.ai/settings/keys` API key path.
- [Anthropic models overview](https://platform.claude.com/docs/en/docs/about-claude/models/overview). Establishes the current Claude lineup, the $10 per million input token price for Claude Fable 5.1 used in the price-ratio check, and that all current models output text only.
- [Exporting your ChatGPT history and data, OpenAI Help Center](https://help.openai.com/en/articles/7260999-how-do-i-export-my-chatgpt-history-and-data). The documented Settings, Data controls, Export data path and the `conversations.json` archive contents.
