Money can no longer buy you more ChatGPT. Build the fallback before your limit runs out

Paying more stopped being the way out of a usage limit

On 2026-09-10 OpenAI paused new sign-ups and upgrades to ChatGPT Pro at 200 dollars a month, the plan it calls Pro 20x. Its help center lists who is caught by it: anyone moving up from Free, Go, Plus, or the 100 dollar Pro plan. Existing 200 dollar subscriptions keep renewing, and the 100 dollar plan is still on sale. The two plans differ only in allowance: 5 times the usage of Plus against 20 times. Fortune reported that OpenAI tied the pause to demand for GPT-6 Astra and said it was working on adding capacity. At the time of writing, on 2026-09-18, no restart date has been published. The practical consequence is new. Until now a usage limit was a pricing question: hit it often enough and you upgraded. If your work already sits on the 100 dollar plan and needs more, that lever is gone for as long as the pause lasts, and it can go the same way at any provider whose compute runs short. The answer is a second platform you can move to the same afternoon, and whether that works depends less on the second model than on where your context lives.

The expensive part of switching is the context you leave behind

Moving a task to another model takes a minute. Moving what the old model knew takes days. Months of work leave context in many places: the memory a chat product keeps about you, the instructions set on a project, the files uploaded to it, and the decisions that only ever existed inside a conversation. Each of those belongs to one product. Even local tools keep memory in their own format and place. Claude Code's auto memory lives under ~/.claude/projects/<project>/memory/, and the documentation says plainly that it is machine-local. It is useful, and no other assistant reads it. Use three tools for a month and you hold three partial memories of the same client, none of them complete. The cost shows up on the day you switch. The new assistant is as capable as the old one and knows nothing: which approach was already rejected and which rule you have given twice. You spend the first hour re-explaining, and the details you forget to mention are the ones it gets wrong. The fix is to keep the context somewhere every assistant can read: plain files, in a folder you own.

One instruction file that three assistants already read

AGENTS.md is the closest thing to a shared format. Codex reads it from ~/.codex for global rules and from every level between the Git root and the current folder, joining them in order and stopping at 32 KiB by default. Claude Code reads it as the project instructions when there is no CLAUDE.md, from version 2.1.277. Where a CLAUDE.md exists, or a session cannot load AGENTS.md directly, a CLAUDE.md whose first line is @AGENTS.md imports it. Gemini CLI reads GEMINI.md by default and accepts AGENTS.md through the context.fileName setting. That gives you one file of standing rules for three tools. It holds what should be true in every session: who you are, the project, the commands, the rules you never want to repeat. Keep it short. Claude's documentation recommends under 200 lines per instruction file, and Codex stops reading at its byte limit, so a bloated file costs context in one tool and gets cut off in the other. The rules file is only half of it. The other half is state: a STATE.md that says where the work is right now, what was decided and why, what was rejected, and the next exact action. AGENTS.md tells every assistant to read it first, and every session ends by rewriting it. With those two files, switching tools stops being a migration. You open the other assistant in the same folder and it starts from the same page.

Split the work by job before a limit splits it for you

A fallback that has never been used is a guess. Route work by job across platforms, so each one already knows the project on the day you need it: the primary takes the heavy daily build, a second platform takes review and research against the same files, and a third stays set up for the day the first two are both short. All of them read the same AGENTS.md and the same STATE.md, so none of them arrives as a stranger. Once a week, run a switch drill. Take one real task, open it in the second assistant from the files alone, and write down what it had to ask you. Every question is a line missing from STATE.md. Watch the usage dashboard as well: when the primary passes 80 percent of its allowance in the middle of the cycle, move review and research off it before the build runs dry. The weak point is upkeep. STATE.md is only as good as the last session that remembered to rewrite it, and most people stop in the second week. That is why I built the Memory Engine. Every 15 minutes it reads the conversation you just had in ChatGPT Codex, Claude Code, Claude Cowork, Gemini or Cursor, keeps what is worth remembering, and files it as plain text in one folder on your own machine. On 2026-09-06 mine had read 667 sessions from Claude Cowork, 250 from Claude Code, 157 from ChatGPT Codex and 10 from Gemini into that one folder. Whichever assistant you open next starts from the same memory, so when a plan stops selling, you open another tool and carry on. If you would rather start by hand, the free Agent Memory Kit on this page sets up the files.

لم يعد المال يشتري لك مزيدًا من ⁨ChatGPT⁩، فجهّز البديل قبل أن تصطدم بالحد

إذا كان عملك اليومي يمر عبر اشتراك ⁨AI⁩ واحد، فأنت تعرف تلك اللحظة: المهمة في منتصفها، والحصة انتهت قبل موعد الـ⁨reset⁩. كان المخرج المعتاد أن تدفع أكثر وتنتقل إلى خطة أعلى، فيعود العمل خلال دقيقة. هذا المخرج أغلقته ⁨OpenAI⁩ أمام كل من لم يصل بعد إلى خطتها الكبرى، ولا أحد يعرف متى يُفتح من جديد.

حين يكون السبب نقصًا في قدرة الحوسبة، يصير الحد سقفًا لا يرفعه المال. والقرار نفسه متاح لأي مزوّد يمر بالضغط نفسه، فالمشكلة لا تخص مشتركي ⁨ChatGPT⁩ وحدهم. سقف من هذا النوع تخطط له قبل أن تبلغه، وأبسط خطة منصة ثانية تنتقل إليها في اليوم ذاته وتكمل ما بدأته.

وهنا يظهر ما لا يدخل عادة في الحساب. المساعد الثاني قادر على العمل، وما يعرفه الأول عن مشروعك يبقى عنده: ما قررته وسببه، وما جربته وتركته، وأين وصلت، وماذا بعد. وكلما طال عملك مع مساعد واحد، كبر ما يعرفه هو ولا يعرفه غيره. نجاح الانتقال يتوقف على مكان هذا السياق: داخل منتج واحد يحتفظ به لنفسه، أو في ملفات على جهازك يقرؤها أي مساعد تفتحه.

الطريقة التي أقترحها تقوم على ملف تعليمات واحد يقرؤه ⁨Codex⁩، ⁨Claude Code⁩ و⁨Gemini CLI⁩، وملف حالة يُعاد كتابته في نهاية كل جلسة، وتوزيع للعمل على ثلاث منصات بحسب نوع المهمة، وتمرين أسبوعي يكشف ما ينقص قبل أن تحتاج إليه. كل ذلك نص عادي في مجلد المشروع، لا يحتاج اشتراكًا إضافيًا ولا أداة خاصة.

في آخر الصفحة مربع النسخ بالإنجليزية، وفيه الملفات، و⁨prompt⁩ لنهاية الجلسة، و⁨prompt⁩ للانتقال، وقاعدة التوزيع، والتمرين الأسبوعي. انسخه إلى مجلد مشروعك كما هو، ثم افتح المجلد في مساعدين مختلفين وقارن النقطة التي يبدأ منها كل واحد.

تابعني على ⁨Instagram⁩:

التسجيل في الدورة: https://halaqa.app/enrollment?course=start-with-ai

الخطوات

Prompt

# PROVIDER-PROOF CONTEXT
# Use this so any AI assistant can pick up your work the day your main plan runs out.

# --- FILES ---
# AGENTS.md (the standing rules, read by every assistant)
# - who you are and how you work
# - the project, its clients or users, and its goal
# - commands, tools and where things live
# - rules you never want to repeat
# - "Before any task, read STATE.md. At the end of every session, rewrite it."
# Keep it under 200 lines. Codex stops reading at 32 KiB by default.
#
# CLAUDE.md (only if a session cannot read AGENTS.md directly)
# @AGENTS.md
#
# Gemini CLI, in ~/.gemini/settings.json:
# { "context": { "fileName": ["AGENTS.md", "GEMINI.md"] } }
#
# STATE.md (where the work is, rewritten every session)
# ## Now
# ## Decided, and why
# ## Tried and rejected, and why
# ## Open questions
# ## Next exact action

# --- END-OF-SESSION PROMPT ---
# Rewrite STATE.md so a different assistant, with no access to this chat,
# could continue tomorrow. Replace what changed instead of appending a diary.
# Record every decision with its reason and every rejected approach with why
# it failed. End with the single next action, specific enough to start without
# asking me anything. Move any rule I had to repeat today into AGENTS.md.

# --- SWITCH PROMPT (the first message in the other assistant) ---
# Read AGENTS.md, then STATE.md. Do not start work yet. Tell me in five lines
# what the project is, where it stands, and what you will do next. Then list
# anything you would still need to ask me.

# --- ROUTING RULE ---
# Primary: the heavy daily build.
# Second platform: review and research, on the same files.
# Standby: set up and tested, for the day both are short.
# When the primary passes 80% of its allowance mid-cycle, move review and
# research off it first.

# --- WEEKLY SWITCH DRILL ---
# Once a week, run one real task in the second assistant with the switch
# prompt. Every question it has to ask is a line missing from STATE.md or
# AGENTS.md. Add the line, then close the drill.