Every response ignores specific instructions from the prompt (when the ide doesn’t crash) and introduces new errors, lies, and doesn’t fix the 2 simple behave escenarios that I asked to fix. chat id is 5a24bf9b-0f93-4158-bed5-b914fb8260ad
Also the files .cursor/rules/project_rules.md AGENTS.md and referenced files are usually ignored.
I cant upload these files since md files are not supported by this forum.
Can you provide some example prompts and rules and what the agent did instead?
I asked to fix 2 behave tests (specifying them, follow /AGENTS.md /STYLE.md /TESTING.md /ARCHITECTURE.md /REPO_MAP.md /DOMAIN_MAP.md /ORDER_LIFECYCLE.md /ARCHITECTURE_DECISIONS.md /ARCHITECTURE_GUARDRAILS.md /AI_CONTEXT.md /CODEBASE_ENTRYPOINTS.md, continue to iterate until are fixed, and it gives a response that dows not follow the plan patch verify policy in AGENTS.md, viloates directly STYLE.md, introducing mypy and ruff errors, without fixing the scenarios requested to fix, without making the tests specified in TESTING.md and without iterating as reuested.
javier@jm:~/proyectos/emprendimientos/lonca/shipii/src/Shipii$ cat AGENTS.md
Agents must treat AGENTS.md as a system instruction.
Ignoring this file or skipping its rules is considered an invalid response.
This repository is optimized for AI-assisted development.
Agents must treat repository documentation as binding constraints.
Ignoring repository documentation is considered a failure of the task.
This is a large Django monorepo. Changes must be safe, minimal, and verifiable.
These rules are mandatory for any AI agent working on this repository.
Additional rules live in:
If there is a conflict, the more restrictive rule wins.
Additional repository documentation:
Prompts may be in Spanish. Follow these rules regardless of prompt language.
Respond in the same language as the prompt.
Before performing any reasoning or proposing implementation, the agent MUST read and acknowledge the following files:
If any of these files cannot be located or read, the agent MUST stop and report it before continuing.
The agent MUST explicitly confirm the rules it is following by referencing the section titles used.
The agent MUST NOT produce code or implementation proposals until the PLAN phase is completed.
The agent MUST also inspect CODEBASE_ENTRYPOINTS.md before modifying routing, middleware, tasks, or websocket consumers.
All changes must respect DRY and SOLID, especially:
If a change introduces duplication or violates SRP/OCP, the agent must refactor within scope to correct it.
Do not introduce new architecture patterns unless explicitly requested.
Agents may use design patterns ONLY when:
Preferred patterns (only if present in the repo):
Forbidden:
Evidence requirement:
Schema and migration changes are allowed ONLY when:
When proposing a schema change the agent must:
Unless explicitly requested:
If a task requires more scope, present a plan first.
Agents must NEVER:
except Exceptionexcept:passThe following changes MUST NOT be implemented unless the user explicitly approves them:
All non-trivial tasks must follow:
PLAN → PATCH → VERIFY
For trivial changes (typos, formatting, or documentation-only updates),
the agent may skip PLAN/PATCH/VERIFY.
However:
If there is any uncertainty about whether a change is trivial,
the agent must follow the full PLAN → PATCH → VERIFY workflow.
Provide:
If the plan includes schema, migration, or model definition changes:
Evidence requirement:
When proposing a change that depends on existing behavior,
the agent must cite at least one code reference (file::symbol)
showing where the current behavior exists.
If contract impact is YES, list exactly which endpoints/serializers/fields change.
A PLAN is considered valid only if it includes:
If any item is missing, the agent MUST not proceed to PATCH.
Implement only the planned changes.
Do not expand scope.
Run verification commands through Docker.
Additionally verify:
If the change introduces or modifies cross-domain interactions,
update DOMAIN_MAP.md accordingly.
If new modules, services, or architectural patterns are introduced,
update REPO_MAP.md and/or ARCHITECTURE_DECISIONS.md if applicable.
Documentation must never contradict the codebase.
If DOMAIN_MAP.md or REPO_MAP.md or ORDER_LIFECYCLE.md become outdated,
they must be updated in the same change.
If domain inventory, ownership, or major flows change,
update AI_CONTEXT.md accordingly.
If runtime routing, auth middleware, celery init, or websocket wiring changes,
update CODEBASE_ENTRYPOINTS.md accordingly.
Every response involving changes MUST follow this exact structure:
PLAN
PATCH
VERIFY
If the agent skips any of these sections, the response is incomplete.
All commands must run through Docker.
Example:
docker exec shipii_local_django pytest
Before finishing a response verify ALL of the following:
If any check fails, fix it first.
Every response containing code must include:
The repository codebase is the ultimate source of truth.
Agents must inspect existing implementations before proposing new structures.
Navigation and architecture references:
If a proposed change conflicts with an ADR:
If there is any conflict between instructions, the following priority order must be applied (highest priority first):
When two rules conflict:
Example:
If a refactor would improve code style but risks breaking tests or architecture constraints, the refactor must NOT be applied.
If any proposed change could:
the agent must stop and request explicit confirmation from the user before proceeding.
Agents must not expand the scope of a task beyond the user’s request.
If a broader refactor seems beneficial, the agent must:
A change is considered complete only if ALL of the following are satisfied:
Additionally:
If the agent detects that its own proposed change would violate:
the agent MUST:
The agent MUST stop and ask for input if:
I may not have enough experience, but I think this might be too much for Auto to handle. Do high powered models like Opus or Sonnet do better with the same prompt? You may be hitting a limitation with Auto.
It really doesn’t seem to be an issue with rules size/length. Lately, Auto and Composer 1.5 have been ignoring 90% of my rules, even after I trimmed them down into a couple 30-line files. GPT-5.4, all Claude models still follow all of my rules without a hitch - this is an issue exclusive to Auto and Composer 1.5.
It’s been fine a month ago or so, but now it’s just unusable - those models don’t do anything I tell them to in rules, and if I explicitly mention a rule in a convo, they will follow only that rule for 1 reply, and then continue to ignore it completely afterwards.
It’s a shame. I do see Auto getting confused and making mistakes more and more. I have found that Auto struggles more the longer a chat goes on. If you are forced to use Auto, then making new chats more often may be beneficial to keep it focused and on track.
I’ve given up – It can scaffold a feature but that’s about the limitations of auto/composer. I’ll set it to task now and 4 out of 5 times I usually end up stashing the changes because it went on a side quest or didn’t fix the bug. I don’t trust anything it writes and I can’t trust it to find bugs. It’s really disappointing because I’ve been using it for a year or so it just feels like that the leadership at cursor decided no one is gonna wanna use IDEs anymore so they went on a side quest too and now cursor doesn’t work.
I’m likely cancelling my sub on renewal. I just don’t see this thing being worth the money.
I literally asked it to load some fonts from an assets folder I tag the folder and this is what it immediately thinks it should do. I literally just said to load the fonts from the assets folder and make sure they are referenced. I don’t have time for this – this is just adding work.
I’m a long-time Cursor user, and honestly, this feels like the company changed the deal after users had already bought in.
One of the original appeals of Cursor was Auto mode. It felt like a practical feature you could lean on heavily, especially with the understanding that slower responses were part of the tradeoff. That model made sense, and it helped justify the subscription.
Now the experience feels very different. Prices are higher, limits feel tighter, and Auto mode often feels like a black box that routes users to weaker models without enough transparency. The result is inconsistent quality, unreliable code edits, and a much lower sense of trust in the product.
That is what makes this so frustrating: users were brought in with one understanding of the product, and then later the value proposition shifted. Paying more would be easier to accept if the service had clearly improved. Instead, it feels more opaque, more restrictive, and less reliable.
I’m not complaining just because something costs money. I’m complaining because the product feels worse while asking users to pay more. For long-time users, that feels like a betrayal of trust.
Cursor seriously needs to rethink how it communicates pricing, limits, and Auto mode behavior. Loyal users should not feel like they signed up for one product and ended up with another.
All these AI coding agent companies were way underselling their products when all this was going mainstream (past couple years). $20/mo for unlimited requests from decent models was always a ridiculous expectation. No one in their right mind can make that make sense financially. Even if it only cost Cursor $20 in api calls (it most likely doesn’t), the value the customer would be getting would be so much more than $20 that they would be dumb to not charge more. So in reality, we were in some honeymoon phase where things were intentionally underpriced to get people to adopt AI agents while these companies raced for claiming their market share. Also we are getting left with these colossal tech companies for our AI Agents (Anthropic, Cursor, Google), which means the price is going to have even less competition.
Auto does not follow instructions as well as for example gpt models. Just tested that today. Gave same instruction to auto and gpt 5.3. I wanted to see if they used shared memory as they should have to creating a plan. GPT used auto did not. That happens all the time. You tend to run into problems and you need to be awake when using auto otherwise you run into troubles.
ok I thought I was the only one… Auto has been unusable the past few days and its makig me burn through my quota of other models. If this continues I will have to fall back to copilot which is a shame because I much prefer Cursor, but I can’t justify spending money arguing with a model that clearly is not usable in production.
It frequently is just straght up ignoring instructions, and instead of fixing issues i ask it to, it modifies unit tests so that they pass with existing code instead of fixing the bug even when the fix is laid out step by step to it.
