Examples

You "I want an image that radiates loneliness."

Pantion "What kind of loneliness? Serene solitude, painful isolation, existential smallness — or something else?"

You "Painful isolation." HARD

Pantion "Is there a person in the image, or should the loneliness be felt indirectly — through an empty space, an abandoned object, a trace of absence?"

You "A trace of absence." HARD

Pantion "Is the absence recent — warmth still lingering — or long ago, with decay setting in?"

You "Long ago." HARD

Pantion "Indoor or outdoor? And should there be clues pointing to a specific person, or should it be universal — so the viewer projects themselves into it?"

You "Outdoor and universal." HARD

Pantion "What kind of landscape? Urban — a neglected bench in an empty park. Nature — an overgrown path, nature swallowing something human. Coast — a washed-up object, an empty pier."

You "An overgrown bench." HARD

Pantion "What season and light? Autumn and dusk are obvious — but maybe you want something unexpected?"

You "Autumn at dawn." FLEX

Pantion "Style — photorealistic, painterly, illustrative? Color palette? And most importantly: what should it NOT be?"

You "Photorealistic. Muted ochre, brown, grey with cold morning light. Definitely not hopeful." HARD

Pantion "How far has nature reclaimed the bench — still clearly recognizable, or barely visible?"

You "Almost completely overgrown, but still recognizable as a bench." FLEX

You "I want to enter a date, time, and reminder text somewhere. When the time comes, I get a WhatsApp message."

Pantion "Where do you enter it?"

You "In WhatsApp itself. I send a message to a contact called Reminder." HARD

Pantion "If you were setting a reminder right now, how would you type it?"

You "Example: '12:15 Piano lesson'. No date means today." HARD

Pantion "When you send a reminder, what do you expect back?"

You "A confirmation showing the interpreted time." HARD

Pantion "What if you don't respond to the confirmation?"

You "The reminder is not scheduled." HARD

Pantion "What if you don't acknowledge the reminder when it fires?"

You "The system resends after 5 minutes, max 3 times." FLEX

Pantion "Is that 3 messages total, or 1 original + 3 retries?"

You "4 total." HARD

Pantion "What if the system is temporarily down when a reminder is due?"

You "A missed reminder is a missed reminder." HARD

Pantion "Can user A see or affect user B's reminders?"

You "No." HARD

Pantion "Timezones. Which 12:00 is '12:00 lunch'?"

You "Amsterdam. Single timezone for MVP." FLEX

Pantion "What if someone enters a time in the past?"

You "Reject it." HARD

Pantion "If someone types '5-6 12:00 lunch' — is that May 6th or June 5th?"

You "June 5th — day first." HARD

Pantion "Can someone have multiple open confirmations at once?"

You "Yes — WhatsApp lets you reply to a specific message. That determines which reminder." HARD

You "I want to enter audio recordings somewhere. The system transcribes them and generates a report in our agency's template format."

Pantion "Where do the recordings come from?"

You "My partner records client conversations at a domestic safety agency. She drops audio files into a shared folder on our home NAS." HARD

Pantion "What audio formats does her recording device produce?"

You "Multiple formats — it varies. M4A, WAV, sometimes MP3." FLEX

Pantion "You mentioned you already have a working transcription app. What does it do, and what model does it use?"

You "It's a drag-and-drop app. She drops an audio file, it transcribes using Whisper large-v3 with speaker diarization — all local, no cloud APIs." HARD

Pantion "The transcription is local. What about the report generation — which model generates the report from the transcript?"

You "A local LLM. Currently Qwen 3 8B running via Ollama. Must also be local." HARD

Pantion "You said 'our agency's template format.' What does that template look like? Is it a fixed structure, or does it vary per case type?"

You "Fixed structure. Header with case metadata, then sections: attendees, key topics discussed, agreed actions, risk indicators, and follow-up date." HARD

Pantion "Where does the case metadata come from — the case name, case number, client identifiers? Does the system need to look these up, or are they embedded in the folder name?"

You "The subfolder name is the case identifier. That's it. No lookup, no database." HARD

Pantion "The report has a 'risk indicators' section. How does the system identify risk? Does the LLM infer them from the transcript, or does the template have a fixed checklist?"

You "The LLM infers them. But it should flag uncertainty — if it's not sure, it marks the indicator as 'requires review' rather than asserting." HARD

Pantion "What should happen when something goes wrong? If an audio file is invalid, or transcription fails, or the report can't be generated — how should she know?"

You "For the MVP, moving it to an error folder is enough." FLEX

Pantion "Should the system run continuously as a background process, or does she start it manually?"

You "It runs on a Mac Mini she doesn't have access to. It must run continuously." HARD

Pantion "What happens after a restart? Should it automatically pick up unprocessed files?"

You "Yes. It scans the inbox on startup and processes anything that hasn't been handled." HARD

Pantion "How does it know what's been handled? A tracking file? A database? Or does moving the file out of the inbox serve as the signal?"

You "Moving it out of the inbox. If it's in the inbox, it hasn't been processed." HARD

Pantion "How does the system know which case a new audio file belongs to?"

You "She creates a subfolder in the inbox with the case name. Every audio file in that subfolder belongs to that case." FLEX

Pantion "When she drops a second audio file in the same subfolder — is that a second conversation within the same case?"

You "Yes. Each audio file is one conversation. A case can have many conversations over weeks." HARD

Pantion "Should the report include a verbatim transcript, or just the structured summary?"

You "Both. The full transcript is saved separately. The report references it but contains only the structured summary." HARD

Pantion "Where does the generated report go?"

You "Into the case folder structure: cases/{case-id}/conversations/{conversation-id}/report.docx, next to the recording and transcript." HARD

Pantion "You said .docx. Is that a firm requirement, or would Markdown or PDF also work?"

You ".docx is required — she opens them in Word and sometimes edits before filing." HARD

Pantion "Given the domain — domestic violence, child protection — what needs to be logged? At minimum: which file was processed, timestamps per step, success or failure with reason, and which model version generated the transcript and report. If you switch models later, you need to trace which model produced which report. This matters for accountability."

You "Yes. Log all of that. Append-only. Every step timestamped. Model version in every log entry." HARD

Pantion "What must the system never do?"

You "Never send data to external services. Never delete files — only move them. Never overwrite a report that already exists." HARD

Pantion "Can she drop multiple files at once? Does order matter?"

You "First in, first out. If she drops five files, they process sequentially." FLEX

Pantion "What is your success criterion for the MVP?"

You "An audio file in the inbox leads to a report in the case folder — without manual intervention." HARD

Pantion "Last question. The NAS path — is it a fixed mount point, or should it be configurable?"

You "Configurable via environment variable. Never hardcoded." HARD

You "Three things changed after real-world testing. First: some recordings are over two hours. Whisper chokes on those."

Pantion "What's the maximum duration Whisper handles reliably?"

You "About 30 minutes. Anything longer needs to be split into chunks first." HARD

Pantion "Should the system split automatically using silence detection, or fixed-length chunks?"

You "Silence detection. Fixed-length might cut mid-sentence." HARD

Pantion "After splitting and transcribing each chunk, the system needs to stitch the transcripts back together. Should it preserve chunk boundaries in the final transcript, or produce one seamless document?"

You "One seamless transcript. The chunks are an implementation detail." HARD

You "Second change: the report needs an overall confidence score. She needs to know at a glance whether the report is reliable or needs careful review."

Pantion "What drives the confidence score? Audio quality? Transcript completeness? Number of uncertain risk indicators?"

You "All three. Low audio quality, many inaudible segments, or uncertain risk flags should lower the score." HARD

Pantion "How is the score presented? Numeric (0-100), categorical (high/medium/low), or traffic light?"

You "Traffic light. Green, orange, red. With a one-line explanation." FLEX

You "Third: the error folder isn't enough. She doesn't check it. She needs a notification."

Pantion "What kind of notification? Email, push notification, a file in a specific location she does check?"

You "A simple text file in the case folder called PROCESSING_ERROR.txt. She opens the case folder to check for the report — if she sees that file instead, she knows something went wrong." HARD

Pantion "Should the error file contain the technical details, or a human-readable summary?"

You "Human-readable. 'Transcription failed — audio quality too low' is enough. Technical details go in the audit log." HARD

You "Compliance says we need a retention policy. Audio recordings must be deletable after the case closes. But the transcripts and reports stay."

Pantion "That conflicts with the existing HARD constraint: 'never delete files.' Do you want to amend that constraint?"

You "Yes. Amend it to: never delete files during processing. After a case is explicitly marked as closed, audio files may be purged. Transcripts and reports are permanent." HARD

Pantion "How is a case marked as closed? A file in the case folder? A command? A separate admin interface?"

You "She creates a file called CLOSED in the case folder. That's the signal." HARD

Pantion "When should the audio purge happen? Immediately when CLOSED appears, or after a retention period?"

You "After 30 days. Compliance wants a grace period." FLEX

Pantion "Should the purge be logged in the audit trail?"

You "Yes. Every purged file, with timestamp and the hash of the original file." HARD

You "Second thing: the agency is updating their report template. When the template changes, she needs to regenerate reports for open cases."

Pantion "Regenerate from the original transcript, or re-transcribe the audio too?"

You "From the transcript. The transcription doesn't change — only the report structure." HARD

Pantion "When she regenerates, should it overwrite the existing report or create a versioned copy?"

You "Versioned. The original report stays. The new one gets a version suffix: report-v2.docx, report-v3.docx." HARD

Pantion "How does she trigger regeneration? Per case, per conversation, or all open cases at once?"

You "All open cases at once. She drops a file called REGENERATE in the root inbox. The system walks all non-closed cases and regenerates." HARD

Pantion "Should regeneration be logged separately from normal processing?"

You "Yes. The audit log should distinguish between 'generated' and 'regenerated', including which template version was used." HARD