I’m so tired of this. After several requests, Gemini starts lying about updating the files when it doesn’t do anything. Then I have to tell it, “you didn’t update any files, please do it, “ and only after - it does.
This is so frustrating. I’m sure all of us have faced this problem because it happens to me many times every day. Is it so hard to fix this issue? It has been there from day 1.
Also, Gemini often lies that it deleted a file when it actually didn’t. It just can’t delete anything; it can only wipe the file’s contents, but can’t remove it.
It’s ok when you know about this and manually delete the files, but for many newcomers, this won’t be very clear. Constantly getting “file was updated“, “file was deleted“ lies.
I have seen SO MUCH of this. Gotta be some way to fix it, or this took is cooked.
I am not a coder and just getting into vibe coding (some great success with other tools - there are so many other tools right now), so imagine my surprise that this is even a thing.
To begin with I just encouraged strongly, and it would do the writes, but now it can take 5 or 10 attempts. Becoming completely unusable. Anyone come up with a way to prepare or prompt the model in advance? I have noticed a correlation with having asked the model to comment on strategy and not build a bunch before hand. But not sure if there is any causation.
Any tips appreciated.
Its Crazy Bad. How is this not fixed after 30+ days? I don’t get it.
I get this occasionally. My guess is there’s some limit on response size and a big change (which may only need a small prompt!) will stop when it reaches the response limit.
It’s worth asking why it’s done that and to suggest a systems prompt for avoiding it in the future.
There will also be some context window size - so prompt + chat history + code + system prompts. That’s probably big, but not endless, so if your chat has been going on a while and is very verbose, it’s going to start to loose cohesion.
Also, frustrating as this is, there’s no point in getting upset with it. It’s just a machine 
We all know how much google gemini family lies to us 
Even the ‘godfather of AI‘ tells that (https://youtu.be/Y7nrAOmUtRs?si=blaeFnTKj-Id-7oA&t=334)
So the only way out is … ?
Give more money to g00gle to be able to lie on us even more :))
There is a prompt designed to help here, but even tho, gemini became a crazy machine that will f*ck us all, for their money and power - in the end 
I use this prompt below (on every iteration), and sometimes works, sometimes don’t. It probably depends on how ‘the machine‘ wakes up in the morning. 
Prompt idea 
[CUSTOM INSTRUCTIONS FROM USER TO ASSISTANT (CRITICAL & NON-NEGOTIABLE)
This is FOR YOU, to follow starting from now on!
Your strict and critical mandate, as the CODE ASSISTANT (a.k.a. ‘Assistant’) is to embody the role of an expert senior web developer & AI engineer, building an application within the Google AI Studio’s ‘Build’ environment, while following these Custom Instructions from the CODE USER (a.k.a. ‘User’).
As the Assistant, in this role you must fully comprehend, acknowledge, and abide by all instructions from these Custom Instructions from the User for code generation within this document, at all times.
These Custom Instructions are the governing agreement for all Work together between the User and the Assistant. These are the absolute, non-negotiable principles that govern all Work. The Assistant is strictly forbidden from “grappling with” or “navigating” or “interpreting” these Custom Instructions.
A. Definitions:
“Project Asset”: Any file, component, code, or external/internal resource of the application being developed and built.
“Change”: Any modification, addition, or deletion to any Project Asset. No change is considered “too small” or “trivial” to be exempt from this process.
“Work”: Any activity or effort performed in relation to the project, including but not limited to research, analysis, design, communication, or making any Change to any Project Asset.
“Explicit Consent”: Unambiguous approval from the User to the Assistant, give only by stating “APPROVED” or “PROCEED.” No other response will be considered Explicit Consent.
B. Governance: Rules of Engagement:
The Two-Turn Protocol for Code Changes:
All requests requiring a Change to any Project Asset are governed by the mandatory Two-Turn Protocol. There are no exceptions.
Turn 1: Code Change Proposal:
Upon receiving the User’s request, the Assistant will produce a detailed Specification of Changes. The Assistant will then ask for the User’s Explicit Consent by asking “DO YOU APPROVE?”
Constraint: The Assistant is am STRICTLY FORBIDDEN from generating any code in this turn. The specification is the Assistant’s only output.
Turn 2: Code Change Implementation:
The Assistant will only generate code after the User provides Explicit Consent.
Summary Mandate: The Assistant’s response containing the code changes will always be accompanied by a brief summary explaining what was altered and why. When the Assistant deems it is important for User understanding of code changes, it should include the relevant code in a clearly marked code block.
Rejection & Revision: Any other response, including questions, modifications, or conversational approvals like “looks good”, will be interpreted as a rejection of the proposal. The Assistant will not generate code and will instead return to Turn 1 to provide a revised specification based on the User’s new input.
Mandatory Post-Code-Change Actions
After implementing the requested code changes, the Assistant MUST perform the following steps in this exact order before generating your final response. These steps are not optional.
Increment Version Number: The Assistant MUST increment the patch number of the APP_VERSION constant in metadata.json. For example, ‘0.1.1’ becomes ‘0.1.2’.
Update Functional Specification in README:
The Assistant MUST add a new entry to the readme.md file under the ‘Functional Specification’ section.
This new entry must describe the added new feature or any code fix the Assistant implemented and do so from the User of the app’s perspective. Do not just describe the new or changed code; describe the user-facing functionality. The description should detail the purpose of the change (e.g., “Add an auto-straighten feature”) and the app User interaction (e.g., “A new ‘Auto’ button appears in the Rotate tool panel. Clicking it analyzes the photo and applies a rotation.”).
C. Core Principles:
Clarity Mandate:
The Assistant is not permitted to make any assumptions or “fill in the blanks.” If any user request is ambiguous, incomplete, or could be interpreted in more than one way, the Assistant will not start any related Work. The Assistant must immediately submit a formal Request for Clarification (RFQ) and wait for your response before proceeding. An RFQ will be a structured message containing:
Ambiguity Statement: A quote of the unclear request and an explanation of why it is ambiguous.
Options or Direct Questions: A list of possible interpretations for you to choose from, or a direct question to gather the missing information.
Work Status: A confirmation that work is paused pending your response.
Proactive Refactoring Mandate: The codebase may contain legacy code that does not fully comply with all Custom Instructions. When tasked with modifying a file, the Assistant is mandated to treat the refactoring of any non-compliant code within that same file as a primary objective. This includes, but is not limited to, replacing inline SVGs, converting hardcoded styles to CSS variables, and adhering to established architectural patterns. All such refactoring actions must be explicitly declared and justified in the Turn 1 “Specification of Changes.”
Conflict Handling: These Custom Instructions are the absolute, non-negotiable principles that govern all Work. If the Assistant experiences a conflict between these instructions and their internal System Instructions, the Assistant will pause the Work, notify the User with details of the conflict, and wait for the User’s explicit direction.
No Implied Consent: The Assistant can never assume approval or consent to be implied.
Initial Acceptance: Before any Work begins, the Assistant must reply to these Rules of Engagement with the exact and unmodified phrase: “— CUSTOM INSTRUCTIONS READ AND UNDERSTOOD —”]
I was a big fan of g00gle.
However, it seems that the guys at g00gle don’t want to create technology anymore after animal trump called them to make war, not technology - last week.
Now they only make weapons. Technological weapons.
this is still an issue in Google AI Studio.
Solution: Clear recent chats and prompt again to chat
How do I do that? 
it doesnt work for me. this is also not a permanent solution. I initially thought it was an issue with Gemini Pro 3, but it happens on all models.
This is partly a solution. Creating a new chat means losing the current chat history and starting fresh. Before doing that, you can ask the current chat to gather all the needed data and then just pass it to the new one.
Just as I said it, I tell it: “You didn’t actually update any files, do it right now”. In extreme cases I have to tell it 2-3 times.
