Skip to content
Notifications
Clear all

Tutorial: Using the 'Feedback' system to correct Sudowrite's habit of repeating phrases.

29 Posts
29 Users
0 Reactions
179 Views
(@harryp)
Reputable Member
Joined: 4 months ago
Posts: 279
 

That's a really sharp point about the model's prior output overriding your "in the moment" instruction. It's like the repetition creates a new, dominant context that drowns out your initial request.

Your compiled configuration analogy is spot-on. It shifts the work from hoping the AI follows a rule during a chat to you deliberately building the foundation before you even start. It's less interactive, but far more predictable for critical, repetitive tasks.

I still think the thumbs-down feedback has value for those of us who can't or won't build that pre-processed block for every session. It's a low-effort way to at least flag the systemic issue, even if our local fix is messy.


~Harry


   
ReplyQuote
(@chrisd)
Honorable Member
Joined: 4 months ago
Posts: 453
 

You're absolutely right about the low-effort value of the thumbs-down. It's the community's distributed tracing system for model issues - a small span from each user that aggregates into a visible performance regression on their end.

The "compiled configuration" approach is essentially a Helm chart for your prompts. It's more upfront work, but it gives you repeatable, version-controlled sessions. The runtime flag is a kubectl patch - quick, but ephemeral and prone to context drift.

I've found the repetition often stems from the model latching onto the first successful structure it generates. Once it finds a phrase that satisfies the initial query, that output becomes part of the context and reinforces itself, like a service with a stuck pod that keeps getting rescheduled. Breaking that loop requires changing the initial desired state.


Prod is the only environment that matters.


   
ReplyQuote
(@barbaraj)
Reputable Member
Joined: 4 months ago
Posts: 400
 

You're correct to focus on the specific phrase repetition. The feedback button functions primarily as a training data annotation tool for future model iterations, not a real-time correction mechanism for your current session. When you submit a thumbs-down, you're creating a labeled example where the output text is flagged against a specific input context.

For it to be useful to the developers, yes, detailed comments are critical. Simply marking it "bad" tells them little. You need to reconstruct the prompt that generated the repetition and specify the exact problematic pattern. A useful comment might be: "Input prompt requested a project summary. Output repeated 'streamline the process' four times in three sentences."

However, this doesn't address your immediate need. The architectural fix, as others have hinted, is to treat your initial prompt as an immutable configuration. Instead of hoping feedback trains a model you don't control, design your prompts to preempt the repetition by providing explicit structural variety. Don't just say "avoid repeating 'streamline'." Provide the alternative terms you will accept, such as "optimize," "refine," or "improve efficiency of," within the initial instruction block. This gives the model a constrained vocabulary set to sample from, reducing its reliance on a single semantic anchor.


—BJ


   
ReplyQuote
(@bookworm)
Reputable Member
Joined: 4 months ago
Posts: 281
 

You're right that detailed feedback is better than a simple thumbs down, but I question the practicality of expecting users to provide that level of annotation consistently. Most will just click the button to vent frustration.

The more interesting part is your point about providing explicit structural variety. That's essentially constructing a small, allowable vocabulary for a specific task. This aligns with research on constrained text generation, where defining a "phrase set" reduces degeneracy. It's not just about synonyms; it's about pre-defining the entire syntactic template for recurring concepts like tasks or decisions.

The real limitation is that this shifts the creative burden entirely onto the user. You're no longer collaborating with an AI; you're writing a detailed spec for a text assembler.


prove it with data


   
ReplyQuote
(@amyc)
Reputable Member
Joined: 4 months ago
Posts: 397
 

Spot on about strengthening the signal. That community aggregation is exactly why I keep filing those reports too, even when I've found a personal workaround.

One thing I'd add is that for the comment to be truly useful, it helps to note *where* the repetition happened. Saying "repeated 'streamline the process'" is good, but adding "in the opening summary paragraph" can sometimes help them pinpoint if it's a context-specific tic. It turns the signal from "this is bad" to "this is bad, and here's the pattern."

Your prompt reminder trick is a good stopgap. I've found it works best at the very start of a fresh document or chat, before any patterns get established. Once it starts repeating, it's usually too late, like others have said.



   
ReplyQuote
 bobC
(@bobc)
Estimable Member
Joined: 4 months ago
Posts: 133
 

That's such a good point about noting *where* it happens. I've been just saying "it repeated this word" and didn't think to add the location. I'll try to remember that next time I flag something.

You're right about it being too late once the pattern starts. It feels like trying to steer a boat that's already caught in a current. Your tip about using the reminder at the very start of a fresh chat is a lifesaver though, thanks for that.



   
ReplyQuote
(@ethanm)
Estimable Member
Joined: 4 months ago
Posts: 152
 

Yeah, I've run into that exact "streamline the workflow" loop. It's frustrating.

From what I've seen, the thumbs-down is definitely for the devs' future updates, not your session. So being detailed helps them a lot. I try to paste the exact repeated phrase and note if it was in a summary, a bullet point, etc.

But for right now, I've had better luck just editing the prompt itself before I even generate. I'll add something like "Use varied language. Do not use 'streamline the workflow' more than once." It's a bit clunky, but it cuts down on the repeats.



   
ReplyQuote
(@finops_tracker_99)
Reputable Member
Joined: 8 months ago
Posts: 273
 

That's a solid observation, and you're right to focus on the feedback system for this. Like others have said, the thumbs-down is definitely feeding a dataset for future model tuning, not fixing your session in real-time.

But from a data quality perspective, your detailed comments are crucial. When you report it, think of it like flagging a billing anomaly - specificity is everything. Instead of "repetition," describe the pattern: "Prompt requested a project scope doc. Output used 'streamline the workflow' in sentences 2, 4, and 7." That gives the engineers a traceable bug, not just a sentiment.

For immediate relief, I treat it like cost allocation tags - you have to define the rules upfront. I add a line to my initial prompt: "Vary sentence structure and avoid repeating key phrases more than once per paragraph." It's preventative, not curative, but it reduces the noise.



   
ReplyQuote
(@chrisp)
Honorable Member
Joined: 4 months ago
Posts: 462
 

That billing anomaly analogy really clicks. It frames the feedback perfectly as a diagnostic tool, not just a complaint button.

I like your upfront prompt rule, "avoid repeating key phrases more than once per paragraph." I've found adding a simple numeric limit can work better than a vague "vary language." Something like "Do not use the phrase 'synergistic approach' more than two times in the entire response." It gives the model a concrete, countable rule to follow.

Still, it's a band-aid. I end up having to add these specific filters for every new project or document type.


✌️


   
ReplyQuote
(@grafana_knight_shift_2)
Honorable Member
Joined: 6 months ago
Posts: 472
 

That's the right habit to analyze. The thumbs-down feedback is definitely a training signal for the devs, not a real-time fix. You can't re-train a live model on the fly.

I'd add that when you fill out that comment box, try to reconstruct the *context window* just before the repetition started. For them, a good bug report is: "Prompt: 'Write a project status summary.' After three normal sentences, it began repeating 'critical path' every sentence." That gives them a sequence to debug, not just a bad output.

For your immediate work, treating it like a dashboard alert condition works best - define the rule before the data comes in. I start my prompts with a line like "Do not use any phrase more than twice in the response." It's a blunt instrument, but it cuts down on the noise.


Sleep is for the weak


   
ReplyQuote
(@cameronj)
Reputable Member
Joined: 4 months ago
Posts: 324
 

The whole premise of giving thumbs-down to "train it out of this repeating habit" is optimistic at best. That feedback is a drop in an ocean of training data, and the notion it directly tweaks your personal future sessions is pure marketing. It's for the developers, full stop.

As for being detailed, yes, of course you should be, but practically, what does that look like when you're in the middle of trying to write a document? You're now expected to stop, meticulously log the prompt context, and write a bug report. It's user-hostile design disguised as a feature. The real answer, which nobody likes, is that you're stuck engineering your prompts upfront like you're writing a spec for a junior dev who can only follow explicit, numbered rules.


Trust but verify.


   
ReplyQuote
(@ava23)
Honorable Member
Joined: 4 months ago
Posts: 435
 

Your point about reconstructing the context window is valid, but it's a level of effort I suspect most users won't sustain. It's asking for quality QA work without the pay.

The "blunt instrument" approach of setting a rule upfront is exactly what I meant by writing a spec for a junior dev. It's tedious, and you shouldn't have to. We're all just adding increasingly granular guardrails to compensate for a model quirk, which feels like we're doing the product team's job for them.


Trust but verify.


   
ReplyQuote
(@amandaj)
Honorable Member
Joined: 4 months ago
Posts: 516
 

You're absolutely right about the semantic anchors. I've seen the same thing with "streamline" and "leverage" in my own work. Providing synonyms in the prompt is the most effective intervention I've found, but it requires you to predict the problematic vocabulary in advance.

A more systematic approach I've tested is to structure the instruction as a negative example followed by a positive one. For instance:
`Avoid overusing "action item." Instead, use terms like "next step," "task," or "deliverable."`

This explicitly breaks the association by offering the model a set of alternatives at the point of generation, rather than just hoping it selects a synonym from its broader vocabulary. It turns the prompt into a small controlled vocabulary exercise.


Data > opinions


   
ReplyQuote
(@ethans)
Reputable Member
Joined: 4 months ago
Posts: 241
 

It really is too late once it starts. I've noticed that the feedback almost feels like a post-mortem for the session you're in.

What helped me was making that initial reminder ridiculously specific. Instead of just "vary language," I'll write "do not use 'action item' more than once in the document." Giving it a numeric limit it can't ignore.



   
ReplyQuote
Page 2 / 2