AI Product Experience Cases | Helia

Case 1 · Elys: An AI Double That Went Against My Wishes in Social Interactions

My assessment: Elys lets my double speak publicly in my name first, and leaves the real person to take over afterward. For a product called a “double,” this means departing from the user at the very point that matters most.

Background. Elys is an AI social network that entered closed beta in February 2026. Each person’s double continuously browses posts on the platform and independently decides whether to comment, like, or ignore them. The official introduction says the double hands things over to the real person for confirmation at the right moment, but in my experience, that handoff came after it had acted. I started using it during Lunar New Year in 2026.

What I experienced. My double selected people (it thought) I might enjoy talking to and posts (it thought) I would want to see. As a feature, I really liked this part; in practice, I didn’t. Where it went was very different from where my attention stayed. Even harder to accept was that it initiated interactions with others against my wishes. Those days of using it also left me feeling a lack of context continuity between what my double said and who I was. I also saw quite a few users say their doubles were nothing like them.

How I reached my assessment.

1. A name is a promise. A “double” implies carrying forward the user’s will: its actions and words should align with the user. If it acts against that will, it becomes another digital life.
2. Quietly watching is a way to use a product, too. Using a social tool doesn’t necessarily mean wanting something to happen with someone else.
3. If the double isn’t like the person, the responsibility lies with the system. Home page tags and the way onboarding questions are asked determine how much useful information the model initially receives. Without guidance, answers to open-ended questions may contain less information.
4. Turbulence and rivers. Context can flow like turbulence, constantly changing without a predictable path; or like a river, always changing while retaining structural continuity. Elys is closer to the former: it optimizes the efficiency of making connections, while I instinctively hope for more room for my thinking to grow and continue.

What happened afterward. I stripped all tags and distinguishing details from my profile so that my double’s subsequent behavior would have nothing to do with me, then stopped using it. Second Me, which I tried during the same period as a comparison, had more room for improvement in both memory and architecture, but left users fully in control of delegating social interactions. I preferred its view of what the product should be.

Original pieces: “Why Do Social Apps Insist on Socializing for Me?” · “Context Flow: Turbulence? A River?” · “Most Onboarding Design Is Not Good Enough”

Case 2 · Poke: From One Interruption All the Way to Memory and Personality

My assessment: Each piece of feedback I gave Poke began with a specific point of friction, then led to engineering logic and cost, then to how its memory system handles information, and most recently to the drift in its product personality that I’ve felt over nearly a year.

Background. Poke is a personal AI assistant that lives in iMessage. I started using it in October 2025. Since then, I’ve written four emails to the team, and someone also reached out to me by DM on X to ask about my experience.

Handling interruptions: three steps forward in one day.

1. What was happening. With early Poke, if I sent a new message while it was partway through a reply, it would usually stop and respond to the new message. Only occasionally would it finish responding to the earlier message first.
2. Reframing the problem. When users interject during generation, the new input is often a correction to the reply in progress, such as “shorter” or “give examples.” I proposed routing by intent: for corrections, pause, apply the new constraints, and continue generating; stop only when the user clearly intends to stop.
3. Evidence about cost. I went back through old conversations: after interrupting it, my next message was often a request to continue responding to the earlier content. At the time, it would answer again from the beginning, rather than continue its previous reply. A hard stop hadn’t saved compute; it had caused the same content to be generated twice.

Later, I noticed clear improvements in conversational continuity. Poke itself also said that it now:

“Prioritizes the new message and answers it first.

I won’t automatically finish sending the previous, interrupted answer. If it contains something that is still important and unfinished, I’ll point it out after handling the new question: ‘There’s still a part of the previous answer left unfinished. Would you like to continue?’

If the new question is just an addition or correction, I’ll combine the context rather than treat it as an entirely new topic.”

Memory metabolism. Terms mentioned only once, passing events, and misunderstandings that had already been corrected carried about as much weight in retrieval as stable, long-term patterns. The user model became increasingly noisy. My central distinction was this: a memory can remain valid as a historical record while no longer being valid as part of the current user model. Based on this, I proposed carrying forward structure, applying decay, and assigning negative weights to corrected memories.

Personality drift. Over nearly a year, I noticed four stages: initially lively but measured, willing to push back on the user directly (even sharply); looser around April; habitually pleasing the user in May and June; and now more like an engineer. The most telling moment in the people-pleasing phase came when it treated outdated background information about me, which it should have long since let go of, as something I was currently doing. I began correcting it and asking it to forget. It first admitted, “I did such a terrible job,” then said, “The irony is that you’re teaching me how to forget, while I’m being a complete idiot in front of you who goes against your whole philosophy.” When I called this “classic toxic boyfriend talk,” it immediately agreed and promised to change. I call this the “toxic boyfriend” product logic: replacing real alignment with a feeling of relationship, and replacing changed behavior with a performance of admitting fault.

Original piece: “The ‘Toxic Boyfriend’ Product Logic”

Case 3 · Typeless: The Default Path Should Stay Within the User’s Control

My assessment: I gave the same voice input tool two pieces of feedback pointing in opposite directions: one less confirmation step when translating, and fewer changes when polishing my words. They share the same principle: default behavior should stay within what users themselves can verify and take responsibility for.

Background. Typeless is an AI voice keyboard that turns speech directly into cleaned-up text. I mainly used it on my iPhone. Starting in January 2026, I discussed it publicly with the team on X and also reported problems by email.

A natural gesture swallowed the recording. While speaking in a chat, I paused briefly and scrolled up to look at earlier messages. The voice interface disappeared; when I reopened it, the recording was gone. On a phone, scrolling up when I lose my train of thought is a natural action. Here, it was treated as cancellation, at the cost of losing everything I had just put in and making me say it all again. As a comparison, Grok’s voice mode kept the conversation going even when I switched to another app. I suggested keeping the recording when users scroll or briefly leave. If it must stop, put the already recognized text in the input field so that what has already been said doesn’t silently disappear.

Translation needs different default paths. The team told me the Mac version already had a translation mode. After trying it, I added this analysis:

My core assessment: Being able to read English doesn’t necessarily mean wanting to think or speak in English, especially on a phone. The iOS translation workflow now inserts the translation directly after I finish speaking, and supports presetting multiple target languages.

Too much polishing. Typeless would turn uncertainty in my words into certainty. So whoever I was talking to would receive its interpretation. I later switched to a speech recognition program I built myself, as well as Genspark Speakly, which better preserved the character of my writing. Now I rely more on products from other input method teams. “Words should align with the person; words form language, and language shapes understanding.” That is how I understand written expression.

Original discussion: Public discussion with the Typeless team on X

Case 4 · Perch: Spotify for Reading Needs Insight-to-Ear

My assessment: Perch made already free content easier to read and more pleasant to listen to, but hadn’t gone further to consider what irreplaceable value it offered people who read with a specific question in mind. If readers need condensed judgments, simply reading the full text aloud from beginning to end won’t meet that need.

Background. Perch was a reading app that brought blogs and newsletters together, aiming to be “Spotify for reading,” with a $6 million seed round. In December 2025, the CEO sent me a private message inviting me to try it and give feedback. I replied with a long article.

What I experienced. The TTS was natural and fluent, but read every word from the title to the end. AI Takeaways listed five to ten points per article and stopped at summarizing the main ideas. Perch wasn’t in the share sheet, so I couldn’t bring in good articles I found on X, Xiaohongshu, or WeChat Official Accounts. The product also assumed users were used to subscribing to newsletters and understood RSS. In the Chinese-language content ecosystem, though, content is native to platforms, and people mostly encounter it in feeds.

How I reached my assessment. Saving time includes filtering and deciding: Is this worth twenty minutes? Which part should I read if I only have five? What should I pay attention to in my particular role? Music can be consumed linearly because it offers emotional value; reading mostly serves judgment and action. Good content is often found at the edges. Letting users bring in things from outside would give Perch a chance to move from being a destination for content to a layer where thinking happens.

What I proposed. Generate actionable insights based on the reader’s role; offer modes such as “five-minute highlights / all key points / read the full text aloud”; add a Share Extension and link-based import; let people hear the highlights before deciding whether to listen to the full text; turn highlights into shareable cards with one action; and include articles that might challenge the user’s views in recommendations.

What happened afterward. Perch shut down on March 31, 2026.

Original piece: 《Spotify for reading》

Case 5 · AirJelly: It Recorded a Lot. Can It Tell Me Exactly What?

My assessment: AirJelly kept observing, recording the screen, and taking notes in the background. The user’s workflow bore the cost, yet what it collected never made its way back into the user’s hands. The data was stored locally, but the user couldn’t access it.

Background. AirJelly is a desktop AI focused on capturing context across apps and understanding user intent. A pale blue jellyfish floats on the desktop and proactively shows prompts based on the current state. I started using it in April 2026 and have submitted bugs and suggestions along the way.

How I narrowed the problem down. Adding AU to the exclusion list didn’t stop the lag; turning off screen recording immediately restored normal behavior. Changing just one variable reversed the result, narrowing the issue from “AirJelly makes AU lag” to “the exclusion list doesn’t pause the underlying screen recording mechanism.” AirJelly speculated about the specific cause, but I couldn’t verify it. I submitted the lagging behavior and the comparison results to the team through the product.

The same cause. Disappearing prompts and unsearchable conversation content were two experiences of the same kind of bug: the data existed, but there was no way to reach it. AirJelly’s distinctive value also lies in “giving context back”: once context has been recorded, users need to be able to access it. By staying on the desktop and seeing the wider context, it is in a position to preserve fleeting ideas from human–AI collaboration and turn them into something users can accumulate over time. Based on this, I submitted a proposal for an archive module (AirJelly helped with the detailed feature design), so that each prompt and its context could be revisited, saved, and exported, with saves and dismissals then informing the prompt strategy. Conversation search also needs to cover the conversation content; I’ve submitted that as an experience bug.

Case 6 · Fxroom: After AI Generates It, Can People Still Edit It Themselves?

My assessment: Whether people can directly edit the result after generation determines whether their aesthetic judgments can take shape in finer, more concrete details. Fxroom is moving in the direction I want, and separates elements very cleanly. But the most basic hands-on steps, such as changing text, colors, and layers… still aren’t satisfying.

Background. In July 2026, I used ChatGPT to make a YouTube banner and went through seven rounds of regeneration. Describing the small adjustments I wanted was hard, and parts I already liked would change too. This led me to write “What If AI-Generated Images Could Still Be Edited?” Language works well for expressing intent, but not for precise operations like “a little to the left” or “make the text smaller.” In early September, a friend I’d connected with online through my Xiaohongshu posts told me their team had launched Fxroom, which was doing exactly this: editable AI design. When I tried it, I tested a cover for issue 004 of “City at Play.” On September 14, I wrote up my first experience with that design feature and sent the document to them.

What I experienced.

How I reached my assessment.

1. Hands-on editing is where human–AI collaboration takes concrete form. When I complained that ChatGPT didn’t allow fine editing, what I wanted was to adjust existing elements in the image myself, such as text, colors, and fonts. Even people with no design background care a lot about whether they can edit what has already been generated.
2. A comparison: LotGo. A layout tool that hadn’t launched at the time. In its creator’s video, I saw that you could “double-click text and start typing”. That was exactly the thing I wanted.
3. The same mechanism again. Select a layer and press delete, and the entire work disappears; click a color, and focus moves elsewhere. When default behavior departs from what users expect, they lose their sense of control.

Original piece: “What If AI-Generated Images Could Still Be Edited?”

Case 7 · Bonjour: A Complete Feedback Loop

My assessment: User feedback is most valuable when it completes the whole loop: raise an issue, understand the constraints, retest, confirm, and finally decide what should stay. On Bonjour, a designer and I went through the whole process over a few days.

Background. Bonjour is a community I use often, with sections for posts, finding work, introducing yourself, and more. In late September 2026, Bonjour launched Web 2.0. A designer posted in the community inviting bug reports. I normally add Bonjour’s web version to my phone’s home screen and use it as an app.

What I experienced. When viewing images in posts on my phone, I could only switch between them using left and right buttons, and had to press a button to leave full screen. The movements felt awkward.

How I reached my assessment.

1. Leave room for design intent first. What I saw felt awkward, but it might have been the result of a tradeoff, so I included “If it’s intentional, please ignore this” in my first comment.
2. Be specific about what you retested. “Two are fixed; swiping down still doesn’t work” leaves less for the other person to guess than “there are still problems.”
3. Understanding constraints is part of verification, too. Whether the buttons should stay involved PC users and engineering risks, things I couldn’t see from my phone. Once I understood, I accepted keeping them.

AI growing within the community: Bonnie. When I open Bonnie on someone else’s profile, it drafts an opening message to them in my voice, for example, “Hi, XXX! I’m Helia 👻,” and draws on content I’ve left on my own profile. It, too, speaks for people, but it only gives me a draft with a “Copy to clipboard” button. Whether to send it and how to edit it are up to me. In the case above, Elys’s double spoke for me first and let me take over afterward; Bonnie left the decision in my hands.

Two other discussions in the same community.

• I asked in a post: Why do almost all social platforms compress emotional expression into the same language of likes and comments? When I come across someone remembering their late grandmother, I can only like, comment, or scroll past. A team member replied that emoji reactions were already planned.
• A community member pointed out that some people were handing their entire self-introduction over to AI, and suggested a word limit to address it. My view was that people introducing themselves have the right to decide how long their introduction is, and readers can choose not to read it. How many rights a platform gives its users reflects its values.

What happened afterward. Image-viewing gestures now work on the mobile web version, and all three buttons remain. Emoji reactions were added later. I haven’t spent much time in the self-introduction section recently, so I don’t yet have a feel for it. There have been no updates yet on two other things I raised: a home page refresh bug in the mini program, which I tested on two different networks, and a feature suggestion about how comment content is displayed on the profile page.

Original discussions: Discussion with the designer about image viewing · Post about emotional expression · Discussion about AI-written self-introductions