I'm experiencing a persistent and repeatable issue with Poe's custom instructions where their efficacy appears to degrade significantly beyond the initial prompts in a conversation. This is causing considerable workflow disruption, as my primary use case relies on consistent, long-form interaction with a specific analytical framework.
I have configured detailed custom instructions designed to shape the assistant's responses for financial operations (FinOps) and cloud cost analysis. These instructions specify a preference for structured data, bullet-point summaries, cost breakdowns, and avoidance of speculative advice. For the first two to three exchanges, the assistant adheres to these parameters perfectly. However, upon continuing the conversation—typically when delving deeper into a cost analysis or asking for a comparative breakdown—the model begins to revert to more generic, verbose, and less structured outputs, effectively ignoring key clauses of the custom instructions.
To illustrate, here is a simplified version of the instruction set I'm using:
```
You are a FinOps analyst. Your responses must:
- Prioritize numerical data and percentages.
- Structure all cost comparisons using bullet points.
- Present pricing in a table format when comparing more than three options.
- Do not provide marketing summaries; focus on total cost of ownership (TCO) and reserved instance coverage strategies.
- When asked for calculations, show the formula used in a code block before presenting the result.
```
A typical failure pattern:
1. **Initial Query:** "Compare 1-year All Upfront vs. No Upfront Reserved Instances for an m5.xlarge in us-east-1."
2. **Initial Response (Adheres to Instructions):** Correctly provides a table with effective hourly rates, calculates break-even point, uses a code block for the formula.
3. **Follow-up Query:** "Now, factor in the potential savings from replacing 30% of this workload with Spot Instances, assuming a 70% interruption rate tolerance."
4. **Degraded Response:** Returns a paragraph-heavy answer discussing the "concept" of Spot savings, fails to structure the blended cost calculation, and omits the explicit formula block, despite the instruction.
This suggests the system prompt or the custom instruction context window is being deprioritized as the conversation token length grows. Has anyone else conducted a systematic review of this behavior? I am particularly interested in:
* Whether this is a known platform limitation with custom instructions and long conversations.
* If there are specific syntactic methods to make instructions more "sticky" (e.g., imperative vs. descriptive language).
* Any official documentation or community findings on the effective context window for custom instructions within a single chat session.
From a cost perspective, this is inefficient as it requires manually re-pasting instructions, starting new sessions frequently, or reformatting outputs, which negates the automation benefit. A quantitative understanding of this limitation would be helpful for planning complex analysis workflows on the platform.
Spreadsheets or it didn't happen.