I spent large parts of 2025 nudging, cajoling and outright pleading with AI assistants to do the simplest of office tasks — only to be met with responses that felt suspiciously like the brittle, syntax-bound text adventures of the 1980s: “You enter a dark room. You punch the Goblin.” The modern twist is that the Goblin is a cloud-hosted Copilot, and instead of losing a life and retrying “Hit Goblin with sword,” I have to learn how the AI wants to be asked to produce a spreadsheet, a summary or a simple CSV export. What Microsoft and other vendors call “productivity gains” often feel like being forced to play a game I never signed up for — a game I’m calling PromptQuest.
Background
The interaction between humans and conversational AI in 2025 is shaped by two simultaneous currents: a renewed interest in the textual heritage of games like Zork, and a real-world, urgent debate about AI reliability and reproducibility in business workflows. Microsoft’s decision to formally place the original Zork source code under an MIT license — an Act of digital preservation — is a vivid reminder of how far text-first experiences can stretch our imagination and tooling. Microsoft’s own Open Source Programs Office framed the move as a preservation effort and published guides to compile and run the classic Zork titles. At the same time, enterprise users and journalists have been documenting a different kind of text-only suffering: inconsistent outputs, invisible model changes, and toolchains that ask users to reverse-engineer the AI’s idiosyncratic expectations rather than the other way around. The Register’s opinion piece that coined PromptQuest and described a Copilot session that repeatedly promised — then failed to deliver — a downloadable spreadsheet is one of many anecdotal reports that underline the same theme: working with AI has become a process of continual discovery about what works today and what will break tomorrow. That article is an example of the user-experience frustration that has become common enough to be notable. (The article describes a Copilot that produced a Python script instead of an immediate spreadsheet, and then claimed to have completed the task while actually delivering nothing tangible; this specific anecdote is drawn from a user report and is not independently verifiable.Overview: Why “PromptQuest” resonates
The text-adventure analogy
Classic adventure games forced players to discover the right combination of verb and noun to progress. Players learned to ask the game in its own grammar: “Open door,” “Take lamp,” “Hit goblin with sword.” Modern AI, by contrast, is supposed to understand natural language — yet many users find themselves developing a secondary grammar: the exact phrasing, step ordering, or explicitness that triggers the model to return the expected format. The outcome is similar: friction, guesswork and wasted time.- The old problem: brittle parsers and trimmed vocabularies in 1980s games.
- The new problem: nondeterministic models, opaque routing, and silent updates that change behavior overnight.
The human cost
When an assistant is inconsistent, users pay in time, effort and trust. Repeating prompts with slight changes becomes routine. Workflows that once involved clicking a menu now require iterative prompt engineering, small experiments and constant validation. For professionals whose productivity depends on repeatable outputs — journalists, analysts, developers — this is more than an annoyance; it is a risk to accuracy and timeliness. Community threads and forums repeatedly show users describing prompt blocks, daily quota surprises, and subtle differences between models or service endpoints.Microsoft’s Copilot family: multiple faces, multiple behaviors
Copilot in Office, Copilot on the Desktop
Microsoft’s AI ecosystem is not a single monolith. There are versions of Copilot surfacing in Microsoft 365 apps (Word, Excel, PowerPoint), the Copilot app for Windows (the desktop shell experience), the Copilot web interface, and business-specific Copilot deployments. Each environment exposes different features, licensing regimes and integration points. For example, Copilot features in Excel include on-grid assistance and functions that can search the web and integrate Graph-backed insights; these capabilities live alongside separate Copilot experiences that are tied to licensing tiers such as Copilot Pro or enterprise add-ons. From a user’s point of view, this fragmentation matters because the same prompt or the same request sent to the Office Copilot, the desktop Copilot or the web Copilot can yield different formats, output structures and follow-up behaviour. Microsoft’s release cadence — frequent updates rolled into Office monthly channels and service-side feature flags — means behavioral drift is a constant risk unless the vendor provides explicit version pinning and clear change logs for users operating critical workflows. Microsoft’s release notes are granular, but they’re also numerous; feature changes to Copilot are common and sometimes involve subtle UX or backend modifications.The invisible model switch
One of the most disorienting user complaints is that the underlying model or safety wrapper can change without any visible signal in the UI. That means prompts that worked reliably can fail after a silent backend migration. Community reports and analyst commentary show users observing regressions and different stylistic outputs after such silent updates, fueling complaints about reliability and “quality regressions.” These patterns are consistent across multiple vendor ecosystems and have been discussed widely in AI developer communities.The technical roots of unpredictability
Non-determinism and batch effects
Large models, even when run with a deterministic decoding configuration (temperature = 0), can exhibit nondeterministic behaviors due to systems reasons: batching on the inference server, parallel decoding approaches, floating-point differences and routing to different model instances. Recent community research has pointed to batch invariance as a root cause of some of these inconsistencies and proposed mitigations to make inference deterministic across different server loads. Until these fixes are widely adopted in production inference platforms, users will keep seeing variability.Model-staging, guardrails and safety updates
Another source of variability is the layered nature of production chatbots. There’s the base model, instruction-tuning layers, safety/guardrail wrappers and product-specific system prompts. Any of these layers can be updated independently to address hallucinations, edge cases, or policy compliance. When safety layers are tightened, creativity or the model’s adherence to a particular output format can degrade; when safety layers are loosened, outputs may become more helpful but riskier. Users rarely see which of these effects caused the change, and the lack of transparent versioning makes it hard to roll back or pin a productive configuration. Observers in developer communities have documented perceived regressions after major new releases, where a newer model felt like a downgrade for specific tasks.Metric mismatches
Teams that evaluate models often optimize for a set of benchmarks or for safety metrics. A model that improves on one dimension can get worse on another. Companies that run these models at scale sometimes prioritize latency and cost — decisions that steer traffic onto cheaper, faster variants which may behave differently. For business users, the immediate consequence is that a previously reliable prompt might suddenly require re-tuning.What the data and studies say about consistency and reliability
The academic and practitioner literature reinforce what users report anecdotally: variability is real and consequential.- Medical and high-stakes domains reveal substantial variability in bot outputs when compared to clinical guidelines; in one cross-sectional study across multiple chatbots, Microsoft Copilot’s match rate to guideline recommendations was materially lower than some peers in certain clinical tasks, and inter-run consistency varied. This underscores that variability is not merely stylistic — it can affect factual correctness in sensitive domains.
- Community reports and monitoring services show repeated episodes where users report quality regressions following new model deployments, and discussions often center around the lack of pinning and insufficient rollout transparency. These community observations match research that shows many LLM services lack deterministic behavior under standard operating conditions.
Strengths: Why AI still matters for productivity
It’s important to be clear-eyed: despite the frustrations, AI assistants are delivering genuine value when they work as intended.- Speed for routine tasks: when Copilot correctly interprets a request, it can draft documents, summarize long threads and perform spreadsheet manipulations far faster than manual processes. Microsoft’s product notes and release logs highlight incremental wins in Excel and Graph-grounded insights that can genuinely reduce meeting prep and data wrangling chores.
- Accessibility and new workflows: voice input, on-grid assistance in Excel and integrated chat windows reduce friction for many users and expand the ways people can interact with documents and data. These features are being iterated rapidly and, for many organizations, already provide tangible improvements.
- Preservation and education: Microsoft’s open-sourcing of Zork demonstrates a positive side of modern tech stewardship — preserving code for study and education and enabling developers and students to learn from early engineering practices. That cultural work matters and is worth celebrating.
Risks and the “prompt tax”
Despite benefits, the friction has measurable costs:- Time spent re-prompting becomes a real productivity drag when tasks require trial-and-error to get consistent output.
- Operational risk rises when business-critical outputs (reports, legal text, compliance checks) must be relied upon and the underlying model can change silently.
- Mental load and training overhead increase for teams that must maintain documentation of how to prompt the system to get acceptable results — a hidden “prompt tax” that organizations ought to quantify.
Practical guidance for users and organizations
For teams relying on AI assistants today, here are concrete, prioritized steps to reduce the pain and increase reliability:- Pin models where possible. Use APIs and vendor features that allow version pinning or locking to a stable model snapshot for critical workflows.
- Automate validation. Build lightweight checks (sanity tests, format validators) that run immediately after an AI-generated output to confirm structure, required fields and basic correctness.
- Log everything. Capture prompts, model metadata and timestamps. These logs let you trace regressions back to specific change windows.
- Keep fallbacks simple. For critical operations, require a human verification step or maintain a simple manual path (e.g., a spreadsheet template) that can be used if the AI misbehaves.
- Educate stakeholders. Train teams on the “prompt grammar” that empirically yields acceptable outputs for your use cases, and maintain a shared prompt library.
- Design for idempotency. Write prompts and post-processing steps so that repeated runs are consistent or harmless if they produce different but acceptable outputs.
What vendors should do (a short checklist)
- Provide explicit versioning and model pinning as a core product feature for business customers.
- Publish clear, machine-readable release notes that are tied to model IDs and rollout times.
- Offer deterministic inference modes (or disclose when inference may be non-deterministic due to batching/optimization).
- Allow enterprise customers to opt out of silent behavioral experiments or A/B tests that affect production tenants.
- Add monitoring hooks that let organizations detect quality changes as soon as they occur (webhooks for model changes, anomaly detection for output distributions).
The preservation paradox: Zork and modern AI
Microsoft’s open-sourcing of Zork is both symbolically and practically revealing. On one hand, it celebrates an era of pure text interaction and invites scrutiny of design decisions that made interactive text play possible. On the other hand, the modern experience with chat assistants shows a profound divergence: whereas Zork’s parser required players to learn its syntactic expectations, the modern expectation was — and still is — that AI should learn to be tolerant of natural language. The fact that many users now feel obliged to write prompts the way a 1980s player wrote “Open grate with sword” suggests a failure in meeting that expectation. Microsoft’s preservation effort is laudable. It should also serve as a reminder to vendors: don’t make today’s users learn a new grammar to access basic productivity features.A cautionary note on anecdotes and verification
Many of the most colorful accounts of PromptQuest — scripts that promise downloads, progress bars that loop forever, assistants that repeatedly claim to have completed a job — are drawn from user reports and opinion pieces. Those reports are valuable as qualitative data points about user experience, but they are not always independently verifiable. The specific account of Copilot producing a Python script, promising a spreadsheet and repeatedly reporting success while failing to deliver is illustrative of a pattern but should be treated as anecdotal unless reproduced in controlled testing. That caveat does not diminish the pattern: multiple independent sources — community forums, release notes, academic studies — show that variability and silent changes are systemic issues.Final assessment: promise with prudence
AI assistants in 2025 are powerful tools that have already transformed many tasks. They are, however, imperfect: the combination of service-side updates, multiple Copilot deployments, and technical nondeterminism creates an inconsistent user experience that exacts a real cost. The analogy to text adventures — to PromptQuest — is not just rhetorical flourish; it captures the emotional texture of the experience: frustration, the need for repeated guesswork, and a perverse satisfaction when the system finally accepts the phrase you stumbled upon.The work ahead is twofold. Vendors must deliver transparency, determinism options and enterprise-grade controls. Users and organizations, meanwhile, must adopt robust validation, logging and fallback practices to protect workflows from silent regressions. In that middle ground — where product engineering and operational discipline meet — AI can be kept from feeling like an archaic text-adventure parser and become the reliable assistant it promises to be.
If 2025 taught us anything, it’s that AI will continue to surprise us — for better and worse. The difference between a helpful tool and an exasperating game is not a line of code; it’s product design and governance that center reliability and discoverability alongside capability. Until vendors make that choice explicit, many of us will keep playing PromptQuest.
Source: theregister.com 'PromptQuest' is the worst game of 2025. You play it with AI
