Macaron playbook

ChatGPT Review: Writing, Research and Plans

ChatGPT is useful for revising writing, code and research questions when you can review its output and control what files you share. A Project can organize chats and supplied material, but it does not grant automatic access to your local folders. Compare the current plan limits, memory settings and data controls before using sensitive work.

By MacaronOperated by MINDAI PTE. LTD.

When Does ChatGPT Fit an Iterative Task?

ChatGPT is worth considering when a task benefits from working with text, files or software over several revisions. Choose it for the actual workflow and account access you need, rather than assuming one plan or interface provides every tool. This account and feature comparison draws on the cited sources; we did not benchmark another assistant.

Public website check. The public Projects guide describes grouping chats, files, and instructions in a project. Official source.

Writing, Coding and Research: Capabilities and Review Requirements

Task	Useful workflow to evaluate	What still needs review
Writing	Supply the source material, audience and requested revision; inspect the returned draft against the source	Unsupported facts, changed meaning and invented quotations
Coding	Work from a defined codebase or example and a concrete change request	Dependencies, permissions and tests in the intended environment
Research	Keep related uploaded files and instructions in a Project, then ask a bounded question	Whether the cited material actually supports each important conclusion

Official Projects documentation describes shared files, instructions and related chats; a project does not automatically provide access to an arbitrary folder on your computer. These jobs can suit iterative work, but none establishes that ChatGPT outperforms another assistant or that a generated program works without checking.

For a bounded household task, use the web interface and the plan actually shown in your account. A text-only brief avoids assuming a file-upload allowance: ask for three dinners for two adults, list permitted ingredients with quantities and units, and request one shopping list. For example:

"Plan three dinners for two adults using the ingredient list below. Keep each meal's ingredient quantities explicit. Group shopping quantities only when the ingredient and unit match. Ask about missing amounts instead of inventing them. Then replace dinner two while keeping dinners one and three unchanged."

Supply the ingredient list after the brief. Compare the first and revised drafts: the selected meal should change, the other meals should stay intact, and shared shopping quantities should reflect the revision. If the list is inconsistent, identify the affected ingredient and ask for a correction; retain a checked text list if a generated file or interface cannot preserve the change. Review ingredient suitability yourself. The three-dinner brief above is a proposed follow-up; the narrower two-dinner web test below was completed.

A two-turn meal plan tested in a temporary chat

An AI agent opened a temporary ChatGPT web chat and supplied fictional inventory for two dinners for two people: 4 eggs, 2 cups of uncooked rice, 1 can of beans and 2 carrots. We asked it to allocate every item, show each day's amounts and verify the total without inventing a can size. It split the inventory evenly: each day had 2 eggs, 1 cup rice, half a can of beans and 1 carrot. Its check table showed zero leftovers.

In the next prompt we made Day 2 egg-free while requiring Day 1 and its amounts to stay unchanged. The revised table kept Day 1 at 2 eggs, 1 cup rice, half a can of beans and 1 carrot; it assigned Day 2 no eggs and left the other three Day 2 amounts unchanged. The new inventory check showed 2 of 4 eggs used and 2 left over, with zero rice, bean or carrot leftovers. It explicitly said that using every original ingredient was now impossible under the new constraints. We checked those four totals against the prompts and the first table.

This is evidence that the visible chat handled one constrained text revision and identified the conflict in that case. It does not measure recipe quality, food safety, nutrition, Projects, file persistence or a shopping-list app. The same ingredient bowls on both days also show why a balanced quantity table is not by itself a useful or varied dinner plan. Test those concerns separately before relying on the result.

A measured limit for news-source lookup

The Tow Center's 2025 test asked ChatGPT Search to identify specific news articles from supplied excerpts. It identified 134 of 200 articles incorrectly and flagged uncertainty in only 15 of 200 responses. The researchers tested article retrieval and citation, not drafting quality, coding or every current ChatGPT model. For research work, a plausible answer or attached link is therefore a starting point: open the cited article, check that its title and passage match the claim, and record the original publisher. We did not repeat the Tow Center's test. Our separate temporary-chat task checked meal allocation, not news citation accuracy.

Hallucinations, Memory and Privacy: The Real Limits

Project organization, memory, model-training controls and scheduled tasks are separate concerns. The published plan comparison lists Projects for Free as well as paid access; they should not be grouped as a Plus-only benefit. Compare actual usage allowances and eligible tools when evaluating paid access, rather than paying for a feature already in your account.

Official documentation directs ChatGPT memory management to Settings → Personalization and distinguishes it from local Codex memories. Memory controls do not by themselves establish how an account's content is used for training. Check the account/workspace data controls separately. Scheduled tasks are documented for eligible plans and surfaces; a task request is not proof that a schedule or external action was created. Exact account limits and the effect of every privacy setting were not tested here.

For a household meal-planning task, try the workflow in your current plan before upgrading: supply a short permitted brief, revise one meal and inspect the changed shopping list. Free Projects already provide organization, so Projects alone are not a paid-plan reason. Consider another allowance or tool only if that specific limitation blocks the job. Our two-turn test below is one narrow workflow observation, not evidence that one assistant is the universal default.

From Chat to Tool: What a Macaron Mini-App Adds

ChatGPT can support an iterative brief. A personal mini-app is useful when you want to return to the same meal records, fields and checklist; compare the actual result rather than assuming either product supplies a feature exclusively.

"Build a weekly meal planner with meal IDs, nonnegative ingredient quantities and units, known pantry amounts and a shopping checklist that recomputes after a meal edit, grouping only the same ingredient and unit and setting buy amount to the larger of zero or need minus pantry, with unknown inputs incomplete and no allergy or nutrition validation."

Open Macaron chat to adapt the request and check the current account offer.

Proposed checks with fictional data:

Need 500 g rice with 200 g in the pantry: show 300 g to buy. A missing pantry amount stays unknown, not zero.
Replace one meal: update its contribution to the shopping list while preserving the other meals and existing checklist state where applicable. If quantities or units are combined incorrectly, correct the rule before using the draft.

This proposed tool was not run. Macaron may process submitted content with third-party AI providers; use permitted, minimal information and review the plan yourself.

Next step

Explore Macaron for a personal tool you can adapt to this task.

Scope and source notes

How this review was checked: Review for Macaron, prepared with AI. An AI agent used a temporary ChatGPT web chat for one fictional meal-allocation task and a constrained revision. Projects, file uploads, memory behavior, scheduled tasks, paid features and the proposed Macaron mini-app were not tested.

Review scope: ChatGPT’s consumer plans, memory and Work workflows. API billing is separate; the comparison uses the same personal meal-planning task.

Sources reviewed: last updated November 5, 2026. Product details may change; confirm current information with the provider.

Source	Supported scope and limitation
Projects and chats \| ChatGPT Learn	Documented workflow or access conditions; not results from this review’s own test
Memories \| ChatGPT Learn	Documented workflow or access conditions; not results from this review’s own test; memory and personalization distinctions
Scheduled tasks \| ChatGPT Learn	Official Scheduled Tasks availability and plan limits; no task or notification was tested
ChatGPT — Pricing	The plan, billing or allowance distinctions in the pricing/limits discussion
AI Search Has a Citation Problem — Tow Center for Digital Journalism	2025 original test of 200 news-article identification prompts per tool; ChatGPT Search misidentified 134 articles. It does not measure writing, coding or current model variants

Corrections: contact@macaron.im.

FAQ

Does ChatGPT have memory?

Yes. Memory is available with account and workspace controls; behavior differs across experiences and settings.

Can ChatGPT produce something beyond text?

Available tools can support files and software as well as text. Check the account and workspace access for the tool you need; a generated artifact still needs review.

Are API charges included in a ChatGPT subscription?

API billing is separate; do not equate an API token price with included subscription tasks.

How should I compare two generated meal planners?

Use the same brief, then check the editing, checklist and review behavior rather than declaring a winner from appearance alone.

What would the Macaron meal-planner example add to a ChatGPT chat?

The proposal uses meal IDs, ingredient quantities and units, pantry amounts and a shopping checklist that should recalculate after one meal changes. The separate ChatGPT text test checked a constrained revision, not persistence or a working Macaron app.

How should I test the proposed Macaron shopping calculation?

With fictional inputs, 500 g of needed rice and 200 g in the pantry should show 300 g to buy. A missing pantry amount must remain unknown rather than becoming zero; the generated tool needs that check before it handles real meal plans.