Evidence-informed Custom Instructions for ChatGPT, updated for the current GPT-5 generation and the 2026 Personalization limits.
Version 4 keeps the strongest idea from v3 — task-specific self-reflection — while removing rigid presentation rituals and adding explicit stopping and verification rules.
Important
OpenAI currently gives Free and Go users up to 1,500 characters, while Plus, Pro, Business, Enterprise, and Education users get up to 5,000 characters. Choose the version that matches your actual limit, not simply “free vs paid.” See the official Custom Instructions FAQ and July 15, 2026 release note.
→ v4 Compact: 1,455 characters
Keeps the core behaviors that fit comfortably inside the smaller limit: task-specific rubric, material-weakness review, stopping rule, evidence preference, adaptive depth, and no mandatory answer ritual.
→ v4 Extended: 3,581 characters
Adds stronger verification rules, context reuse, ambiguity handling, rewriting preservation, validation guidance, and task-adaptive formatting while leaving room for future edits.
v3 introduced a useful pattern: ask the model to silently build a task-specific rubric and use it to improve the answer. That idea is not just prompt folklore. OpenAI's GPT-5 prompting material explicitly describes using a private 5–7 category rubric for complex tasks. See the OpenAI GPT-5 Prompting Guide and its source in the public openai/openai-cookbook repository.
The newer GPT-5.6 guidance, however, pushes in another direction at the same time: simplify prompts, keep success criteria and stopping conditions, remove redundant scaffolding, and validate with tools when possible. See OpenAI's live GPT-5.6 prompting best practices.
v4 tries to combine both lessons.
The rubric stays because it gives reflection a concrete target. The exact ≥98/100 threshold from v3 is replaced with a practical criterion: fix material weaknesses, then stop when further iteration is unlikely to materially improve the result.
Why: self-refinement can help, but repeated self-correction without new evidence is not guaranteed to improve an already-correct answer.
- Self-Refine: Iterative Refinement with Self-Feedback
- Large Language Models Cannot Self-Correct Reasoning Yet
Keep going until solved with a best score sounds ambitious but has no natural stopping condition. v4 explicitly stops when the important rubric dimensions are satisfied and more rewriting is unlikely to help.
That matches current GPT-5.6 guidance, which recommends clear success criteria and stop rules rather than unnecessary process scaffolding.
v4 still allows the model to silently apply the standards of the relevant expert domain, but it no longer forces a visible persona, fictional prestige, or award into every new chat.
Expert personas can affect behavior and style, but evidence that they reliably increase factual accuracy is mixed; irrelevant persona detail can also hurt performance. See When “A Helpful Assistant” Is Not Really Helpful: Personas in System Prompts.
A TL;DR is useful for a long research answer and pointless for a two-sentence factual reply. Step-by-step decomposition is valuable for some reasoning and troubleshooting tasks but should not be mandatory presentation syntax for every answer.
v4 tells the model to use these structures when they improve the result, rather than because the template always demands them.
When current sources, files, tools, calculations, tests, or other validation are available, v4 tells the model to use them. A second internal opinion from the same model is weaker evidence than an actual check.
The extended version makes this explicit for research, calculations, code, data analysis, and tool-based work.
The v4 prompts contain no contributor-specific profession, language preference, personality preset, memory contents, connected services, subscription-only tools, or personal writing style. They are intended as general-purpose defaults. Features are referenced only conditionally — “when available.”
OpenAI's current documentation says:
- Free and Go: up to 1,500 characters
- Plus, Pro, Business, Enterprise, Education: up to 5,000 characters
Source: ChatGPT Custom Instructions — OpenAI Help Center.
The 5,000-character expansion for the higher tiers was announced on July 15, 2026: ChatGPT Release Notes.
v4 is evidence-informed, not yet benchmark-proven.
The previous v3 MMLU-PRO run is preserved in the v3 archive, including its reported 70.20% overall score and the author's note that a TL;DR-related grader bug caused some answers to be misclassified.
A fair next test should compare baseline vs v3 vs v4 Compact vs v4 Extended using:
- the same model and snapshot;
- the same reasoning effort;
- the same benchmark sample;
- the same decoding/settings;
- the same grader and parsing logic;
- raw outputs retained for inspection.
Until that test is run, v4 should be treated as a reasoned proposal rather than a claimed numerical improvement.
- Open ChatGPT → Settings → Personalization.
- Open Custom Instructions.
- Choose v4 Compact or v4 Extended based on your plan's character limit.
- Copy only the prompt inside the code block and save it.
- Start a new chat; Custom Instruction changes apply to future conversations.
This repository and the original v1–v3 approach were created by Denis Shiryaev (@DenisSergeevitch); the original project is chatgpt-custom-instructions.
I used the v3-style instructions for a long time and found the core self-reflection/rubric idea useful. Rather than replacing it wholesale, I eventually went back through the current OpenAI guidance and the research around self-refinement, self-correction, and persona prompting to see which parts still made sense. v4 is the result: preserve the strong core, remove arbitrary or overly rigid pieces, and make the prompt fit today's ChatGPT limits.
Thanks to Denis for publishing the original prompt, benchmarks, and iterations openly — they are what made this review possible.
- GPT-5.6 prompting best practices
- GPT-5 Prompting Guide — OpenAI Cookbook
- GPT-5 Prompting Guide — source on GitHub
- ChatGPT Custom Instructions FAQ
- ChatGPT Release Notes
- Self-Refine: Iterative Refinement with Self-Feedback
- Large Language Models Cannot Self-Correct Reasoning Yet
- When “A Helpful Assistant” Is Not Really Helpful: Personas in System Prompts
Feel free to use and modify these instructions for your own use.