← sources

Heavy Codex Use - Hit Limits Easily Still Never Finished (Claude alt?)

source: reddit (reddit:ChatGPT)published 2026-10-07 17:35ingested 2026-10-08 02:36

retained text

I run a small business and use Google Sheets + Apps Script for internal operations. Keeping my industry and business numbers private, but these are real systems employees depend on, not practice projects.

I use ChatGPT Pro plus a separate Work Premium seat. I keep hitting usage limits, waiting for resets, and then burning through more usage trying to finish or repair the same systems. I’m considering the $500 plan, but before spending more, I want to understand whether my workflow is actually efficient.

More usage would help me keep working. What I’m questioning is how much of that work produces lasting reliability.

**What I’m building:**

* **Master Feed:** Imports reports from an existing business platform, normalizes the data, and supplies several downstream spreadsheets.
* **Aging dashboard:** Shows unpaid balances and how overdue they are. It needs scheduled refreshes and clear indicators when the information is stale.
* **Accounts receivable workstation:** Lets staff select an account, review balances and invoices, and manage collection follow-ups.
* **Delivery hub:** Tracks delivery assignments, status, outstanding pickups, and deposits.
* **Commission ledger:** Connects payments and credits to invoices and employees so commissions reflect money actually collected, including returns and reversals.
* **Returns tracker:** Matches returned items to purchase history, identifies duplicates, tracks supplier eligibility, prepares return requests, and archives completed records.

These systems share data, so a successful import doesn’t necessarily mean every downstream dashboard or invoice allocation is current.

**My workflow:**

I explain the problem to ChatGPT, usually using Sol/Astra for analysis and detailed Codex prompts. Codex handles implementation, troubleshooting, and verification. For difficult work involving live financial spreadsheets, I use stronger models and higher reasoning settings.

Then I bring the results, logs, checkpoints, and unfinished items back to ChatGPT to prepare the next prompt.

That has produced useful functionality. But it also creates a lot of back-and-forth, and I’m constantly wondering whether the next prompt is moving us toward completion or starting another expensive investigation.

**What’s frustrating:**

I keep dealing with stale refreshes, Apps Script timeouts, trigger problems, dropdown failures, duplicate account keys, and downstream sheets that haven’t fully adopted the newest data. Some repairs pass their immediate checks, but another scheduled run or normal employee workflow exposes a problem.

Other times, code is tested locally or in staging, but permissions or tool-access blocks prevent deployment. That’s a different problem from bad code, but from my perspective I’ve still spent time and usage without getting the working system I need.

I’m also concerned about repeated audits, rereading checkpoints, and checking things that were already verified. I understand why financial systems need validation. I don’t know how to tell when the agent is being appropriately careful versus spending a lot of usage without resolving the main bottleneck.

**My target is practical functionality.**

I want roughly 80–90% of the intended workflows usable and dependable so employees can get their work done. I can finish secondary features and polish later. I’m not saying incorrect balances or lost records are acceptable I’m saying I need the essential workflows finished before we keep expanding the scope.

Right now, I feel like I’m spending too much time maintaining the tools that were supposed to save me time.

**For people using these tools on real operational systems:**

1. Is ChatGPT planning → detailed Sol prompt → Codex execution an efficient workflow, or am I adding unnecessary handoffs?
2. How do you measure agent efficiency beyond “it worked for an hour” or “the tests passed”? What tells you a repair will actually hold up?
3. Would Claude/Claude Code help as an independent reviewer or debugger alongside Codex, with only one tool making production changes?
4. Would Cursor materially improve this kind of Apps Script project through better code navigation and editing, or would I still face the same Google permissions, deployment, and runtime problems?
5. Would you spend more on Codex usage, add a second provider, or first change the architecture and development process?

I’ve tried Claude before, and my prompts haven’t translated as well, so switching would have its own learning curve. It didn't do basic things for me that I know it can do so I canceled my 20$ Claude plan but instead of upping to $500 Pro I'm considering staying on $200 and using a high end Claude subscription.

I’m looking for feedback from people maintaining interconnected Sheets/Apps Script systems over time. What actually reduced recurring failures, usage consumption, and the amount of supervision you had to provide?

document fetched 2026-10-08 00:20 · canonical https://reddit.com/r/ChatGPT/comments/1x02uzb/heavy_codex_use_hit_limits_easily_still_never