ChatGPT Custom Instructions by DenisSergeevitch: A Rubric-Based Prompt Pack

What the chatgpt-custom-instructions repository actually ships

This is a prompt collection, not software. The repository holds a README, two earlier versions (v1.md and v2.md), and three image files: v2_mmlu.jpeg, v3_graph.png and v3_radar.png. There is no package manifest, no build step, no test suite, and no installable artifact. The README describes the contents as "My optimized custom instructions for ChatGPT and Operator that improve performance." The default branch is main, and the last push was on 2026-03-14.

The intended user is someone who already uses ChatGPT's custom instructions field and wants a starting template instead of writing one from scratch. The README's own framing is personal: these are the author's prompts, published so others can copy them. That matters for how you read the rest of the repository. There is no support commitment, no changelog beyond the versioned Markdown files, and no compatibility matrix for ChatGPT clients other than the note that the instructions work with Voice Mode.

If you were hoping for a tool that injects instructions through an API or a browser extension, this is the wrong repository. If you want a compact block of text to paste into a settings field and then edit, it is exactly the right size.

How the v3 prompt works: a private rubric, then a rewrite loop

The mechanism is stated directly in the prompt. The model is told to spend time constructing a rubric from a role's point of view, with 5 to 7 categories, and to keep that rubric hidden from the user. It then scores its own draft against the rubric and rewrites if any dimension falls short of a high threshold. In the README's wording, the model should "internally think and iterate on the best (≥98 out of 100 score) possible solution" and start again when the response does not hit top marks across all categories.

The answer itself is shaped by a second block. On the first chat message the model assigns itself a real-world expert role, for example a named PhD in a specific topic with a local award, and then answers in that role. Output follows a fixed structure: a one-line TL;DR (skipped for rewriting tasks) followed by a step-by-step answer with concrete detail. The README also lists style defaults that read as constraints rather than features: no tables unless requested, and no unsolicited "what to do next" suggestions.

What is interesting here is what is not in the prompt. There is no retrieval step, no tool call, no external state. The whole loop happens inside the model's reasoning. That keeps the instructions short enough to paste, but it also means the rubric is unobservable. You cannot inspect the categories the model chose, and you cannot tell from the outside whether a rewrite actually happened or whether the model simply produced a fluent answer and stopped.

Installing the template and running a first prompt

There is nothing to install in the conventional sense. The README's "How to Apply" section gives four steps: open ChatGPT, go to Settings, select Personalization, and paste the instructions into the "What traits should ChatGPT have?" field. The prompt text itself lives in a fenced block in the README and in v2.md and v1.md for the older revisions.

A practical way to get the text without retyping it is to clone the repository and read the file:

git clone https://github.com/DenisSergeevitch/chatgpt-custom-instructions
cd chatgpt-custom-instructions
sed -n '/^## Instructions/,/^## How to Apply/p' README.md

The sed range prints from the Instructions heading up to the How to Apply heading, which is the block you want to copy. If you prefer the previous revision, the same content is in v2.md.

Once the text is in the traits field, a first test that exercises the mechanism is to ask for something with a checkable answer, for example a short explanation of a concept in a domain you know well. The README's expected shape is a role line, a TL;DR, then a step-by-step answer, with no table and no closing list of next steps unless you asked for one. If you see a table or a "what to do next" block unprompted, the style defaults are not taking effect on your client and the template is not doing what it claims.

The README notes compatibility with Voice Mode, which is worth knowing because voice replies will not show the role line or the TL;DR formatting in the same way.

The MMLU-PRO numbers and the evaluation bug the README admits

The README reports a v3 run on MMLU-PRO with 8,447 correct out of 12,032, or 70.20% overall. The per-domain table is more informative than the headline. Math is the strongest domain at 86.75%, Chemistry follows at 79.68%, and Physics at 78.60%. Law is the weakest at 46.78%, with History at 57.48% and Engineering at 61.61%. Those spreads are large enough that a single overall number hides most of the story.

The README is also candid about the setup and about a defect. The run used GPT-5 Nano with medium reasoning "to keep costs low", not the larger GPT-5 models the prompt is written for. And an evaluation bug is disclosed: a first-line TL;DR in the template caused a subset of answers to be misclassified by the grader. The author states the v3 prompt still outperformed the baseline despite the caveat and that the run will be repeated. No baseline numbers appear in the README, so the size of that improvement cannot be checked from the repository alone.

Treat the table as a single unreplicated run on a small model with a known grading defect. That is not a reason to dismiss it, but it is a reason not to quote 70.20% as a property of the prompt. The prompt is model-agnostic in principle; the measurement is not.

Where this template breaks down

The hidden rubric is the main limitation. Because the model is told never to show the rubric, you have no way to audit whether the categories are sensible for your task or whether the self-scoring is meaningful. A rubric invented silently for a legal question and a rubric invented silently for a math problem may look nothing alike, and the prompt gives you no handle to steer either one.

The role assignment is the second soft spot. The prompt instructs the model to claim a real-world expert identity with a prestigious local award. That is a stylistic device, and it can produce confident framing on topics where the model's underlying knowledge is thin. The Law domain score of 46.78% is a useful reminder that the persona does not add knowledge.

The third issue is fit. The template is tuned for long-form explanatory answers. If your work is short classification, extraction, or structured output, the TL;DR line and the step-by-step format are overhead, and the no-tables default actively works against tabular extraction. The README does not document how to disable individual rules other than editing the text, and it does not describe what happens when these instructions conflict with a project's own instructions or with memory.

Compared with writing your own instructions or using a generator

The obvious alternative is to write your own custom instructions. The trade-off is control versus speed. Writing your own takes an afternoon of iteration, but you end up with rules you understand and can debug. Pasting this template takes a minute, but you inherit a rubric mechanism you cannot inspect and style defaults that may not match your work. If you go the template route, the repository's own version history is the useful part: v1.md and v2.md let you see what changed between revisions, including the removal of what the README calls non-working hacks such as promising the model a reward or claiming the author has no fingers.

A second alternative is a prompt generator, the kind of tool that asks you a few questions and emits a custom instructions block. The difference in approach is that a generator produces text tailored to your stated preferences, while this repository ships one fixed, opinionated block. Neither is better in the abstract. A generator gives you a starting point closer to your use case; this repository gives you a starting point closer to a specific prompting methodology, the rubric-and-rewrite loop, that a generic generator is unlikely to produce.

A third option is to skip custom instructions entirely and put the rules in a project or a saved prompt. That keeps the instructions scoped to one kind of work instead of applying to every conversation, which is the right call if you only need the structure for a narrow task.

Maintenance, licensing and what the repository does not say

The last push was on 2026-03-14, so the repository has not been touched for several months. There is no release, no tag, and no issue tracker activity described in the README. Updating means re-reading the README, comparing it against v2.md, and re-pasting the block into ChatGPT's settings, because there is no mechanism to sync the text automatically. If ChatGPT changes the Personalization field, the README's four-step apply procedure is the only guidance available, and it will not have been updated.

The licence section says only: "Feel free to use and modify these instructions for your own use." There is no SPDX identifier and no licence file in the repository listing, which means the terms are informal rather than a standard open source licence. For personal use that is unlikely to matter. For redistribution inside a company or a product, the absence of a formal licence is a real question, and it is worth asking before you build on it. Nothing here is legal advice; the point is simply that the permission text is one sentence long and does not address attribution or commercial redistribution.

Editorial conclusion

Adopt it if you want a short, readable custom instructions template you can paste into ChatGPT's Personalization settings and edit yourself; the v3 text is compact and the repository keeps v1 and v2 for comparison. Do not adopt it if you need something you can install, version-pin, or run in CI, because there is no package, CLI or API here, only Markdown. Before pasting, open v3 in the repository, read the answering_rules block in full, and confirm the ChatGPT settings path the README gives (Settings, then Personalization, then the traits field) still matches your client, since the README does not document what happens when a client moves or renames that field.

Frequently asked questions

Official sources

1. DenisSergeevitch/chatgpt-custom-instructions on GitHub
2. Issues
3. Project website
4. README