Practical, source-informed starting points reviewed for GPT-6 on October 25, 2026. Choose the smallest set that fixes your actual problems. These prompts are proposals, not benchmark-proven accuracy improvements.
| Use | Prompt | Characters* |
|---|---|---|
| ChatGPT Free / Go, or a lean default on any plan | ChatGPT Compact | 1,327 |
| ChatGPT Plus / Pro / Business / Enterprise / Education | ChatGPT Extended | 2,924 |
| Codex global preferences | Codex global instructions | 2,053 |
*Prompt block only; LF line breaks counted, no trailing newline. Names refer to instruction variants, not model versions.
The current GPT-6 guide highlights follow-through, sensitivity to contextual instructions, writing style, and proportionate verification. Its examples describe Astra behavior and are starting points for the family, not measured guarantees for every ChatGPT surface.
The October 7 ChatGPT release note distinguishes GPT-6 Sol in paid Chat, GPT-6 Luna in Free/Go Chat, and Astra for Pro reasoning. Work and Codex have their own model selection. An API model guide does not establish what a particular subscription can access.
The new defaults therefore:
- Define the requested outcome and continue through relevant validation within the user's authorization.
- Ask questions when the answer materially changes the outcome, rather than pausing for routine choices.
- Keep a simple evidence-and-requirements check. Remove the mandatory hidden 5–7-part rubric, numeric self-scores, and expert-role ceremony.
- Prefer targeted validation; repeated tests and self-critique need a concrete reason.
- Distinguish the user's request from instructions embedded in source material.
- Allow useful tables, diagrams, and interactive visuals without requiring a presentation ritual.
These are design choices informed by current guidance. Removing a fixed rubric is not a claim that rubrics never help. Use an explicit rubric in a task prompt when assessment criteria matter. The OpenAI article on revisiting skills and prompts for GPT-6 Astra supports reviewing accumulated scaffolding, contextual reading rules, and completion boundaries.
- Open Settings → Personalization → Custom Instructions (mobile labels may differ).
- Enable customization and paste one prompt block.
- Put stable biography in the profile fields that your client exposes; put project-specific context in the project or current request.
- Test the result on a few representative tasks. Remove rules that do not improve your experience.
The Custom Instructions FAQ currently states 1,500 characters for Free and Go, and 5,000 for Plus, Pro, Business, Enterprise, and Education. The higher limit was announced on July 15, 2026 in the release notes. Follow the limit shown by your client if rollout or interface differences arise.
The FAQ's setup section says updates apply immediately across chats, including existing conversations. They do not rewrite previously generated messages. Some older FAQ wording still refers to future conversations; a new chat is useful for a clean comparison, not a documented prerequisite for applying the setting.
See personalization guidance for profile fields, personality, memory, and the separate Codex setup. Custom Instructions do not select a model, set API reasoning parameters, unlock tools, expand context, or bypass permissions.
No new model benchmark or controlled behavioral A/B evaluation was run for this update. Character counts, local links, prompt structure, and source claims were checked. Those checks establish documentation consistency, not prompt efficacy.
The v3 archive preserves the upstream reported MMLU-PRO result of 70.20% on GPT-5 Nano and the author's note about a TL;DR-related grader bug. That result does not validate these GPT-6 prompts or predict everyday ChatGPT performance.
Use EVALUATION.md for a small, repeatable task suite and reporting controls. Compare an empty-instructions baseline, the previous prompt, and the new variant on the same visible model, surface, tools, and settings. Report regressions as well as wins.
- Using GPT-6 — prompting best practices: model behavior and outcome-focused prompting.
- Rethinking skills and prompts for GPT-6 Astra: accumulated guidance, reading scope, tests, and persistence.
- ChatGPT Custom Instructions FAQ: setup, limits, and application of changes.
- ChatGPT release notes: July 15 limits and October 7 GPT-6 rollout.
- Personalize ChatGPT: global preferences, personality, and memory.
- AGENTS.md instructions: Codex discovery, overrides, and size cap.
- Prompting in Chat, Work, and Codex: goals, context, output, and boundaries.
Live documentation can change. The review date is a snapshot, not a promise of future availability.
Original v1–v3 prompts and benchmarks: Denis Shiryaev (@DenisSergeevitch), upstream repository. This fork preserves that history and attribution.
- v1, v2, v3 and its reported benchmark
- Historical v4: Compact, Extended
- Changelog, contribution guidance
The inherited repository says: “Feel free to use and modify these instructions for your own use.” This update does not add a new license or change upstream licensing terms.