Dr. Markus Schmidberger

Ask first

• Which channels should it watch, and how often? (default your main video channel, hourly)
• Which comment types may it reply to on its own, and which must wait for you? (default reply to praise and content suggestions; hold business inquiries, complaints, and links)
• What tone should replies use? (default two sentences max, warm, no emojis or hashtags)
• Where should harvested audience requests go? (default a backlog table with vote counts)
• What is the maximum number of replies per hour? (default 15)

Needs

• Read and reply access to your channel's comment feed. Connecting: "It reads public comments on your videos and posts replies only to categories you approved."
• A table of already-processed comment IDs and a backlog table for audience ideas.
• A short voice guide with three sample replies you wrote yourself.
• If your assistant can only chat: paste a batch of recent comments once a day and ask for replies to copy back plus a list of viewer requests.

Build

1. Read: each hour, fetch comments posted since the last stored watermark across videos from the past 30 days. First run reads 7 days back, capped at 100 comments.
2. Decide (classify): sort each comment into one of five buckets: idea request, positive feedback, technical question, business inquiry, or spam/hostile. Keep idea requests when they name a concrete subject ("do a 1990 off-road truck next"). Drop vague praise ("nice") from the idea backlog.
3. Decide (dedupe ideas): compare each new idea against existing backlog rows; if a match exists, increment its request count instead of adding a duplicate row.
4. Decide (draft reply): for positive feedback, idea requests, and simple questions, draft a one-sentence reply in the host voice. Hold business inquiries, links, or negative comments in a review list. Cap at 15 auto-replies per hourly run.
5. Deliver: post approved categories, append new or incremented ideas to the backlog table, and send a summary message only when a business inquiry is held. Fallback: save all replies as unsent drafts on a daily review page.
6. Memory: store every processed comment ID so no comment is ever answered twice. About 720 runs a month, 1 to 5 model calls per run.

Check

• Run with posting turned off on the last 20 channel comments: verify spam is ignored, a concrete viewer suggestion is added to the backlog table, and a duplicate suggestion increments the counter.
• Run a second time on the same window: zero new replies or rows.
• Quiet run: an hour with no new comments logs one line and sends nothing.

In my version

• Runs hourly on Google Cloud Run via Cloud Scheduler using Gemini 3.8 Flash, reading YouTube comments and logging car requests into Firestore for the Showrunner agent.

Ask first

• How many days without a call, meeting, task, or note makes an account "stalled"? (default 14 days)
• Which deal stages count as active? (default everything except Closed Won and Closed Lost)
• How many items should appear in the daily brief at most? (default top 5 overdue or stalled items)
• May it send follow-up emails automatically? (default never; generate subject and body drafts only)

Needs

• Read access to your CRM's accounts, opportunities, tasks, and activity log. Connecting: "It reads deal dates and activity timestamps to spot quiet accounts and draft emails; it never sends emails or alters deal stages without your action."
• If your assistant can only chat: export a CSV of open deals with their last-contact date once a week, paste it into chat, and ask for the top five stalled accounts and follow-up drafts.

Build

1. Read: query open opportunities, overdue tasks, and all activities logged across the last 14 days. Cache the brief for 1 hour so reopening the panel does not re-run the model.
2. Decide (overdue & decayed): flag any open opportunity whose close date is in the past, any task past its due date, and any active account with zero linked activities in the 14-day window. Rank by deal value and days overdue, keeping the top 5.
3. Decide (draft follow-up): for each flagged account, read the last 3 notes or emails on that contact and draft a short follow-up email referencing the last conversation topic.
4. Deliver: render the daily brief in the assistant panel with one-click buttons to view or edit each unsent email draft. Fallback: deliver the brief as a morning email to yourself.

Check

• Set a test account's last activity date to 20 days ago and run the brief: confirm it appears under Stalled Accounts with a follow-up draft citing the last note.
• Log a new note on that test account and refresh: confirm the account drops off the stalled list.
• Quiet run: when all active accounts have touchpoints within 14 days and no deals are overdue, it displays "All active accounts touched within 14 days."

In my version

• Built with Kiro specs as the Proactive CRM Agent inside our EspoCRM deployment, combining a deterministic activity analyzer with Gemini to produce the daily brief and follow-up drafts.

Ask first

• What 3 topics or audience problems do you want to speak about? (default leadership lessons, technical tradeoffs, and founder stories)
• How long should each voice session last, and how many follow-up questions should the interviewer ask? (default 8 minutes, 3 to 4 probing follow-ups)
• Words and formatting habits never to use? (default no rhetorical "Let's dive in", no buzzwords, no emoji bullets)

Needs

• A voice recording or live speech-to-text interface.
• A voice DNA document containing 5 to 10 of your best past posts and a banned-phrases list.
• If your assistant can only chat: turn on voice dictation in your chat app, ask it to interview you one question at a time for 8 minutes, and then have it draft 3 posts from your transcript.

Build

1. Read (interview): start with one seed topic and ask a concrete opening question ("Tell me about a time this week when..."). Listen to the spoken answer and generate a follow-up question that pushes for a specific number, example, or failure mode. Stop at 8 minutes.
2. Decide (extract): pull verbatim quotes, numbers, and story beats from the transcript so the draft uses only claims actually spoken in the interview.
3. Decide (draft): write 3 distinct post drafts using the voice DNA rules, each tied back to the exact timestamp in the transcript. Reject any draft containing banned filler phrases.
4. Deliver: save the drafts to the weekly content calendar for human review before anything is posted. Fallback: email the 3 drafts and transcript to yourself.

Check

• Record a 2-minute test answer containing one specific number and run the drafter: confirm every claim in the resulting post traces back to your transcript and zero banned phrases appear.
• Quiet run: if the microphone disconnects or the recording is under 60 seconds, it alerts you immediately and generates no fake drafts.

In my version

• Built with Kiro specs inside my voice-to-LinkedIn web app, feeding drafts and transcript provenance into my weekly content calendar.

Ask first

• Which email and calendar accounts are in scope? (default your primary work inbox and main calendar)
• When should the morning brief arrive, and on which messaging app? (default 7:30am local time by text or WhatsApp)
• Which senders or meeting types always count as urgent? (default investors, customers, board members, and same-day travel)
• May the assistant move internal meetings or send replies without asking? (default never; propose the change and wait for a yes text)

Needs

• Read and draft access to email, plus read and write access to your calendar. Connecting: "It reads new mail and calendar slots to spot clashes and write drafts; it never sends mail or moves external meetings without your confirmation."
• A messaging channel where you can reply in a few words.
• If your assistant can only chat: paste your day's calendar and unread email subjects each morning and ask for clashes, top three priorities, and reply drafts.

Build

1. Read: each morning at the scheduled time, pull unread emails from the last 24 hours and calendar events for the next 48 hours.
2. Decide (triage): group incoming mail into Needs a decision today, Waiting on others, and Low priority. Flag any calendar overlap, back-to-back calls with zero buffer before an external meeting, or unbooked meetings agreed to in email threads.
3. Decide (drafts): for each item in Needs a decision, prepare a short reply draft or two concrete alternative time slots from free calendar windows.
4. Deliver: send one concise message listing only the clashes and decisions needed, numbered so the person can reply "1 yes, 2 move to Fri". Wait for explicit confirmation before sending any email or moving any calendar invite. Fallback: save the brief as a pinned note at the top of the inbox.
5. Memory: record handled thread IDs and resolved clashes so they are not repeated the next morning. About 30 morning runs a month plus on-demand text turns.

Check

• Create an overlapping test event on tomorrow's calendar and send yourself an email asking to meet: run with delivery off and confirm both the clash and two valid free slots appear in the brief.
• Quiet run: a morning with no clashes and only newsletters sends a two-line "All clear" schedule summary.

In my version

• Instinct connected to my inbox and calendar, texted throughout the day to triage mail, coordinate schedule changes, and clear daily admin.

Ask first

• How many short videos should run per day, and at what release times? (default 1 to 2 per day at midday and evening)
• How long should the private review buffer be before a video goes public? (default 24 to 48 hours)
• What is the monthly generation budget cap, and how many retries are allowed per failed shot? (default $300/month, max 3 retries per shot)
• Where should the preview link and veto alert arrive? (default Telegram or text message)

Needs

• A backlog of episode concepts and visual prompt rules for your niche.
• Any text-to-image model, image-to-video model (or Ken Burns pan-and-zoom over still images as a low-cost fallback), a vision model for QA, and FFmpeg for assembly.
• Upload credentials for your video channels. Connecting: "It uploads compiled videos as private with a future publish time; nothing goes public immediately."
• If your assistant can only chat: ask it to write the 10-shot storyboard, image prompts, and motion prompts for today's concept, then paste keyframes back into chat to check for visual glitches before editing.

Build

1. Read: on schedule, check the monthly cost ledger against the budget cap. Pull the next unproduced concept from the backlog.
2. Decide (storyboard): generate a 10-shot storyboard with continuity anchors so each shot builds on the previous stage.
3. Read & Decide (visual QA): generate each clip and extract keyframes. A vision model scores each clip from 0 to 1 on physical plausibility, subject stability, and prompt adherence. Keep clips scoring 0.75 or higher. Retry failed clips up to 3 times with targeted corrective prompts; if an experimental concept still fails, fall back to a proven baseline template.
4. Deliver (compile & schedule): normalize audio, stitch the 10 clips with FFmpeg, set synthetic media disclosure flags, and upload as PRIVATE scheduled 24 to 48 hours ahead. Send a message with the preview link, QA summary, cost, and scheduled hour. If you take no action during the veto window, the platform publishes it automatically; one tap holds it private.
5. Memory: log every shot receipt to the cost ledger and save the episode manifest. About 30 to 60 runs a month.

Check

• Run the pipeline in dry-run mode on one concept with uploads disabled: confirm failed keyframes trigger bounded retries (max 3), audio is loudness-normalized, and the cost ledger increments accurately.
• Quiet run: if monthly spend reaches the cap, the runner halts before generating video and sends a budget alert.

In my version

• Runs on Google Cloud Run Jobs via Cloud Scheduler for CarWash96 using Gemini 4 Argon (Director), Nano Banana 2.1 (base frames), Veo 3.1 (video), and Gemini 3.8 Flash (Visual Critic), uploading to YouTube and TikTok with a Telegram preview to me.

Ask first

• What breathing cadence should the visual cuts follow? (default 16 seconds total: four 4-second clips matching inhale, hold, exhale, hold)
• What visual flaws should Stage-A keyframe QA reject before generating video? (default wrong body stance, broken equipment geometry, or wrong camera angle)
• What is the monthly budget cap for this channel? (default $150/month)

Needs

• Discipline blueprints defining the 4 camera perspectives (entry, glide, apex, seamless loop return).
• An image generator for the 4 anchor keyframes, a vision model for two-stage QA, a video generator (or slow-motion pan fallback), and FFmpeg for 432Hz audio mixing.
• Isolated channel upload credentials so each channel publishes to its own account and schedule.
• If your assistant can only chat: ask it to write 4 camera-continuous prompts timed to a 4-4-4-4 breathing cycle, generate the 4 still images first, and paste them into chat for a geometry check before animating.

Build

1. Read: pull the next wave discipline and lighting blueprint and verify the channel's monthly spend is below its cap.
2. Decide (Stage-A keyframe gate): generate the 4 reference-conditioned 9:16 keyframes first. Audit each still image with a vision model for stance and geometry rules. Regenerate any failed still image before calling the video model, saving roughly 80% of wasted video spend.
3. Decide (Stage-B motion gate): animate each approved keyframe into a 4.0-second slow-motion clip and audit mid-clip frames to reject vertical bouncing or unnatural physics.
4. Deliver: mix phase-locked acoustic breathing cues at 0s, 4s, 8s, and 12s over a 432Hz ambient bed, stitch the 16-second loop with FFmpeg, upload as PRIVATE for the channel's staggered release slot, and send a Telegram preview.

Check

• Feed a deliberately flawed keyframe into Stage-A with video generation off: confirm the critic rejects the still image and makes zero video API calls.
• Quiet run: if the channel's monthly cap is reached, the job exits cleanly and logs a budget notice.

In my version

• Powers Channel #2 (OceanFlow96) inside my multi-channel studio monorepo, auditing Nano Banana 2.1 keyframes with Stage-A critic rules before running Veo 3.1 Fast and 432Hz audio assembly.

Ask first

• Where do your daily notes live? (default a local folder of markdown files such as an Obsidian vault)
• How many days back should each scan look? (default 3 days, or since the last scan timestamp)
• What topics or content pillars are you writing about? (default 3 core themes you define once)
• Which notes or folders must never be scanned? (default private journals, client folders, and credentials)

Needs

• Read access to your local notes directory, excluding private subfolders.
• A list of your already-published post titles so the scanner does not suggest repeats.
• A session-start hook in your IDE or assistant, or a daily script.
• If your assistant can only chat: paste your rough notes from the last two days into chat and ask for the two strongest post hooks that do not repeat your recent posts.

Build

1. Read: when a workspace session starts, find markdown files modified in the last 72 hours outside the excluded folders.
2. Decide (filter): skip daily checklists, empty stubs under 50 words, and notes already marked as drafted. Score the remaining notes for concrete lessons, counter-intuitive findings, or real architecture tradeoffs. Keep: a note detailing why a production deployment failed and how it was fixed. Drop: generic meeting agenda bullets.
3. Decide (angle): match the top 3 notes to your content pillars and write a one-line hook and one-line takeaway for each, checking against the published archive to avoid duplicates.
4. Deliver: print the 3 candidates at the top of the session and append new ones to the ideas backlog file so the voice interview routine can pull them as seed questions. Fallback: write the candidates to a single markdown file on your desktop.

Check

• Save a test note describing a technical lesson learned, then trigger the session hook: confirm it surfaces the note with a hook line and ignores notes inside excluded folders.
• Quiet run: if no notes changed in the last 72 hours, it outputs one line ("No new note candidates") and touches nothing.

In my version

• A Kiro SessionStart hook scanning my Obsidian vault whenever I open my LinkedIn strategy workspace in Kiro and writing candidates to my backlog.

Ask first

• Which days of the week and what time do you publish? (default Monday, Wednesday, Friday at 1:00pm local)
• How are your draft folders organized? (default one folder per ISO week with files named by weekday)
• What engagement actions do you want on the checklist? (default reply to comments from the last post and engage for 30 minutes after publishing)

Needs

• A local folder of scheduled post drafts and a parked-ideas file for backup drafts.
• A date-checking script or rule that verifies weekday names against the real calendar year.
• If your assistant can only chat: paste your weekly drafts folder list and ask what is scheduled for today and whether any dates are mismatched.

Build

1. Read: check today's real calendar date and ISO week number, then open the matching weekly folder and the posting schedule rules.
2. Decide (today's plan): if today is a posting day, locate the draft for today's weekday. If the slot is empty, pick the top finished draft from the parked file. If today is not a posting day, list only comment follow-ups and backlog drafting.
3. Decide (date guard on save): whenever any markdown file in the content folders is created or saved, parse every date and weekday mentioned in the filename, frontmatter, and body against the system calendar. Flag any off-by-one mismatch immediately.
4. Deliver: display the ready-to-paste post text followed by the numbered engagement checklist. Fallback: print the checklist in the terminal.

Check

• Save a draft file named "wed_test.md" dated on a Thursday: confirm the on-save hook catches the weekday mismatch immediately.
• Trigger the daily routine on a posting day: confirm it outputs the exact current week's post ready to copy.

In my version

• Two Kiro hooks in my strategy workspace: a manual daily routine hook that checks the current ISO week and builds my checklist, and a PostFileSave hook that validates dates on every saved markdown draft.

Ask first

• What day and time should the weekly episode be generated, and when should it go public? (default generate Thursday morning for Friday evening release, leaving a 36-hour review window)
• What is the target duration and aspect ratio? (default 8 to 12 minutes, 16:9 landscape)
• Should on-screen people ever speak to the camera? (default never; keep all voiceover strictly over hands-on action B-roll so there is never uncanny lip-sync)

Needs

• A library of stories or topics to adapt into conversational scripts.
• Any two-voice text-to-speech engine, a B-roll video generator (or a folder of reusable workshop B-roll clips as a low-cost fallback), a vision QA model, and FFmpeg.
• YouTube upload access configured to schedule as private first.
• If your assistant can only chat: paste a raw story or outline and ask it to write a two-host conversational script paired with a shot-by-shot B-roll list that avoids talking heads.

Build

1. Read: every Thursday morning, select the next story from the library and load the host persona rules.
2. Decide (script & visual decoupling): write an 8-to-12-minute dialogue script with natural banter. Pair every scene with silent tactile action B-roll prompts; strip any prompt that asks for a person speaking on camera.
3. Read & Decide (QA): synthesize the voiceover tracks, generate or pull the matching B-roll clips, and check keyframes with a vision model to reject any clip showing open mouths or visual glitches.
4. Deliver: mix audio with sample-accurate alignment and -16 LUFS loudness normalization, compile the 16:9 video, upload as PRIVATE scheduled for Friday evening, and send a Telegram notification with the preview link. Fallback: save the compiled MP4 to cloud storage and send a download link.

Check

• Run a 60-second test chapter with uploads off: verify the voiceover aligns cleanly across cuts, loudness measures -16 LUFS, and the vision critic rejects any frame with a talking face.
• Quiet run: if Friday's release slot already holds a scheduled long-form video, the script skips generation and alerts you.

In my version

• Triggered every Thursday at 06:00 UTC via Cloud Scheduler and Windows Task Scheduler, combining Gemini 4 Argon scripting, text-to-speech audio, Veo 3.1 action B-roll, and a 36-hour Telegram veto window.

Ask first

• Which grocery store or delivery service do you use, and what is your preferred delivery window? (default your primary local supermarket, Friday afternoon)
• What are the 15 to 25 staple items your household buys almost every week? (default seeded from your last 3 order receipts)
• What is the budget ceiling where it should warn you before checkout? (default $150)
• May it complete payment on its own? (default no; load the cart and ask for confirmation before placing the order)

Needs

• A running list of staples and a way to text items as you run out in the kitchen.
• Access to your grocery store account or past order receipts. Connecting: "It adds items to your cart and checks delivery slots; it never checks out without your confirmation."
• If your assistant can only chat: text items to a single chat thread all week, then on Thursday ask for a deduplicated shopping list sorted by store aisle.

Build

1. Read: throughout the week, capture any message like "add olive oil and berries to groceries" into the open weekly list.
2. Decide (merge): once a week on order day, combine the week's texted items with the recurring staples list. Deduplicate quantities and skip long-shelf-life items (such as laundry detergent) if bought in the last 3 weeks.
3. Deliver: add the merged list to the online grocery cart, select the preferred delivery window, and send a text showing the added items, any out-of-stock substitutions, and the total price. Wait for your reply before finalizing the order. Fallback: send the grouped checklist to your phone before you shop.

Check

• Text two items twice during the week and trigger a dry run: confirm they appear once in the summary alongside your staples and nothing is purchased without confirmation.
• Quiet run: if you tell it you are traveling that week, it holds the staples cart and orders nothing.

In my version

• Managed through Instinct by text: I drop items in chat whenever we run low, and it coordinates the weekly grocery order for my sign-off.

Ask first

• How many hours after publishing should a video be evaluated? (default 48 hours after public release)
• What retention or view threshold triggers a rescue? (default views below 50% of your 14-day channel median or 3-second hold below 65%)
• May it swap live titles and thumbnails automatically, or wait for approval? (default propose the swap and wait for one-command approval)

Needs

• Read access to your channel analytics (views, average percentage viewed, and retention curve) and the prompt manifest for each video.
• Write access to update video metadata once you approve.
• If your assistant can only chat: paste the titles, view counts, and average view durations of your last 10 videos into chat and ask which two underperformed, why, and for three replacement titles.

Build

1. Read: pull performance metrics for videos published between 2 and 14 days ago, alongside the exact shot prompts used to generate each video.
2. Decide (rescue): flag any video under the 50% median view threshold whose average view duration is still healthy (meaning the content holds viewers, but the packaging failed) or where the first 3 seconds dropped sharply. Generate 2 replacement hook-first titles and a high-contrast thumbnail frame.
3. Decide (learn): correlate shot attributes (subject type, camera angle, opening action) with completion rate across the last 20 videos. Increase selection weight for top-quartile patterns and down-weight bottom-quartile patterns.
4. Deliver: output the rescue candidates and updated prompt weights; apply title and thumbnail updates only after approval. Fallback: write a weekly markdown report.

Check

• Feed a mock analytics table where one video has high completion (80%) but low views (30% of median): confirm it flags that video for a title/thumbnail rescue rather than a content rewrite.
• Quiet run: when all recent uploads track above the median baseline, no metadata changes are proposed.

In my version

• Runs via my rescue and market-feedback scripts in my media studio repo, pulling YouTube Analytics data to repackage underperforming uploads and update prompt weights for the Creative Director.

Ask first

• How many entities or accounts are you keeping books for? (default one company account; ask which company before filing if more than one)
• What is your tax jurisdiction and VAT/GST treatment for software and foreign vendors? (default local VAT rules and reverse-charge tagging for cross-border services)
• Where do raw bank CSVs and invoice PDFs get dropped? (default a local inbox folder per company)

Needs

• Exported bank statement CSVs and a folder of PDF/image invoices and bills.
• A rules file mapping recurring vendor keywords to chart-of-accounts categories and tax codes.
• A filing calendar of statutory return dates.
• If your assistant can only chat: upload an invoice PDF or paste anonymized bank CSV lines and ask it to extract vendor, net, tax, gross, and account category as a table.

Build

1. Read: scan the inbox folder for new bank CSV exports and unprocessed invoice or bill PDFs. If multiple companies exist and a document's buyer name is ambiguous, stop and ask before filing.
2. Decide (classify bank rows): run the keyword classifier across every CSV row to assign category and tax treatment; flag unknown vendors for review rather than guessing.
3. Decide (extract invoices): read each PDF or image to pull invoice number, date, supplier, line items, net amount, VAT rate, and total into structured JSON, validating that net plus tax equals gross.
4. Decide (reconcile): match extracted invoices against bank debits by amount and date window (within 7 days); list any bank debit over the receipt threshold that still lacks a matching document.
5. Deliver: write the validated JSON records to the company ledger folder and output a summary showing matched totals, missing receipts, and the next filing deadline.

Check

• Drop a sample invoice PDF where net + VAT does not equal the printed total: confirm the extractor flags the math discrepancy instead of silently saving it.
• Quiet run: re-running on an already reconciled month produces zero duplicate ledger entries.

In my version

• Built in Kiro with custom steering files for accounting rules and invoice extraction, managing multi-tenant books and filing calendars for my Irish limited companies.

Ask first

• Which calendar keywords or attendee counts mark a meeting as high-stakes? (default board, investor, keynote, interview, negotiation, or 6+ external attendees)
• How many minutes before the meeting should the nudge arrive? (default 10 minutes)
• How long should the breathing cycle last? (default 3 minutes of 4-4-4-4 box breathing)
• What is the maximum number of nudges per day? (default 3, spaced at least 2 hours apart)

Needs

• Read-only access to the device calendar. Connecting: "It checks event titles, start times, and attendee counts on your device; calendar text never leaves your phone."
• A local notification timer and a simple log of completed sessions.
• If your assistant can only chat: paste today's calendar each morning and ask it to pick the two highest-pressure meetings and set phone alarms 10 minutes before each.

Build

1. Read: every morning and whenever the calendar updates, scan events for the next 24 hours.
2. Decide (high-stakes): flag an event when its title matches the high-stakes word list, when an external domain executive is attending, or when three back-to-back meetings leave under 15 minutes before an external call. Keep: "Series A Partner Pitch". Drop: "Weekly internal standup".
3. Decide (throttle): if more than 3 events qualify in one day, keep the highest-ranked 3 and space nudges at least 2 hours apart. Skip if a session was already completed within the last 90 minutes.
4. Deliver: fire a local notification 10 minutes before start time that launches a 3-minute paced breathing timer, asks for a 1-to-10 tension rating before and after, and saves the delta. Fallback: a calendar reminder set 10 minutes early.

Check

• Create a test calendar event titled "Board Review" starting in 15 minutes and run the check: verify a single notification is scheduled 10 minutes before start.
• Add a second test event 30 minutes later: verify the 2-hour cooldown suppresses the second nudge.
• Quiet run: a day with only internal 1-on-1s schedules zero notifications.

In my version

• Built in Kiro using spec-driven development as my Android box-breathing app, reading the phone calendar on-device and tracking pre/post state-shift cards.

Ask first

• How many rejected clips should accumulate before compiling a reel? (default 10 clips)
• Which failed clips are funny vs unusable? (default keep physics glitches and extra tools; drop blank, dark, or corrupted frames)
• What time slot should bonus reels take so they never collide with regular uploads? (default an evening bonus slot 6 hours after the main daily upload)

Needs

• A storage folder ("vault") where your QA script moves rejected clips along with a one-line JSON note explaining why the clip failed.
• FFmpeg to stitch clips and overlay short captions.
• If your assistant can only chat: keep a folder of funny failed AI generations and ask your assistant for punchy 4-word on-screen captions before stitching them in your editor.

Build

1. Read: whenever the main video pipeline finishes a run, move any clip rejected by the visual critic into the blooper vault if it is visually intact (not blank or corrupted) and failed due to an absurd physics or object glitch.
2. Decide (trigger): count uncompiled clips in the vault. When the count hits 10, select the 10 clips and turn each critic failure note into a 3-to-5-word self-aware caption ("Wait... two spray guns?").
3. Deliver: stitch the 10 clips into a 45-second vertical video with loudness-normalized audio, upload as PRIVATE scheduled for the next open evening bonus slot, and send a preview link to your phone. Mark the 10 vault clips as used.

Check

• Place 9 sample clips in the vault and run the hook: confirm nothing compiles. Add a 10th clip and run again: confirm one 45-second compilation is built and the 10 clips are marked used.
• Quiet run: runs with fewer than 10 vault clips exit in under a second.

In my version

• My blooper vault service hooked into the cloud showrunner, recycling Veo 3.1 clips rejected by the Gemini 3.8 Flash critic into zero-cost bonus Shorts at 18:30.

Ask first

• Every how many episodes should a compilation be built? (default every 10 episodes)
• How should episodes be ordered inside the compilation? (default put the highest-retention episode first as the hook, then chronological or alternating pace)
• Which day and time should compilations go live? (default Sunday afternoon at 15:00 so it never collides with weekday uploads)

Needs

• Archive of master video files and their performance metrics.
• FFmpeg for chapter transitions and audio normalization.
• Channel upload credentials for private scheduled publishing.
• If your assistant can only chat: paste the titles, durations, and retention scores of your last 10 videos and ask for the best chapter order, compilation title, and timestamped description.

Build

1. Read: after each episode finishes, check the total episode counter. Trigger when the counter is a multiple of 10.
2. Decide (rank & order): pull the last 10 completed episodes and sort by retention rate to place the strongest episode in slot 1, followed by a balanced sequence of the remaining 9. Generate chapter timestamps for the description.
3. Deliver: concatenate the master files with smooth audio crossfades and -16 LUFS normalization, generate a composite thumbnail, upload as PRIVATE scheduled for the upcoming Sunday at 15:00, and send a preview link to your phone.

Check

• Run the compiler on 3 short test clips with upload disabled: verify chapter timestamps in the generated description match the exact cut points in the output MP4.
• Quiet run: episodes 1 through 9 exit immediately without running FFmpeg.

In my version

• My milestone compilation script in the media studio pipeline, triggered on every 10th episode milestone and scheduled for Sunday 15:00 Ireland time.

Ask first

• Who is traveling, and what are your airline, seat, and hotel preferences? (default direct flights where possible, aisle seats, quiet 4-star boutique hotels with flexible cancellation)
• What is the total trip budget ceiling? (default $1,500 for a short break, $4,000 for a week)
• May it pay for non-refundable bookings without asking? (default never; present 3 options and wait for explicit sign-off before charging anything)

Needs

• Read access to your calendar to check clear date windows and school or work holidays.
• Saved travel preferences (home airport, loyalty numbers, room preferences).
• If your assistant can only chat: tell it your dates, home airport, budget, and vibe, and ask for 3 complete flight-plus-hotel pairings with cancellation rules to book yourself.

Build

1. Read: when texted a trip request ("find us 4 warm days in November"), check the calendar for open windows and load saved travel preferences.
2. Decide (shortlist): search flights and accommodations matching the window, filter out multi-stop itineraries and non-refundable rooms unless requested, and rank the top 3 complete pairings by travel time and value.
3. Deliver (options): text the 3 options with door-to-door travel times, total cost, and cancellation deadlines.
4. Deliver (book after approval): once the person replies with their choice, complete the booking or hold, forward confirmation numbers, and block travel and stay dates on the calendar.

Check

• Ask for a weekend trip on dates where your calendar already has a committed event: confirm it flags the conflict before proposing flights.
• Run a shortlist request and stop before replying: confirm zero payment methods are charged and zero calendar blocks are written until you reply with an option number.
• Quiet run: if no flights or hotels meet the budget and direct-flight rules, it reports the constraint that broke and books nothing.

In my version

• Handled by texting Instinct with dates and constraints; it researches routes and hotels, sends me the shortlist, and coordinates the bookings once I approve.

Ask first

• What clip length range works best for your channels? (default 15 to 30 seconds)
• How many clips should be extracted from a typical 30-to-60-minute video? (default top 5 moments)
• Should clips publish immediately or go into a scheduled queue for review? (default schedule across the next 5 days with a review step)

Needs

• A tool to download the source video and its timestamped transcript.
• A language/vision model to score transcript segments and FFmpeg to crop to 9:16 and burn in subtitles.
• Optional: a social scheduler API to queue the finished clips.
• If your assistant can only chat: paste the timestamped transcript of your talk and ask it to pick the 5 punchiest 20-second segments with exact start/end timestamps and social captions.

Build

1. Read: download the source video and its word-level or line-level subtitle file from the provided link.
2. Decide (select hooks): scan the transcript for self-contained 15-to-30-second segments that open with a strong claim and end on a clean resolution without mid-sentence cuts. Score and keep the top 5 non-overlapping segments.
3. Deliver (cut & schedule): cut each segment with FFmpeg, crop to 9:16 vertical framing, burn in high-contrast word captions, generate a social post caption for each, and upload to the social scheduler queue. Fallback: save the 5 MP4 files and captions in a local output folder.

Check

• Run on a 10-minute video with social posting off: verify all 5 output clips are between 15 and 30 seconds, start and end on complete sentences, and have synchronized captions.

In my version

• My Python clip-cutter pipeline powered by Gemini 3.8 Flash and FFmpeg, scheduling vertical clips to Instagram Reels, TikTok, and LinkedIn via GoHighLevel.

Ask first

• What is the narrative arc across the 4 segments (0-8s, 8-15s, 15-22s, 22-29s)? (default problem setup, tension peak, product interaction, calm resolution)
• What aspect ratio and resolution do you need? (default 9:16 for social reels or 16:9 for landing pages)
• Does your video model support native clip extension? (default yes; if not, use last-frame chaining with FFmpeg)

Needs

• One high-resolution reference image anchoring the visual style and character.
• A JSON campaign file defining the initial 8-second prompt and three 7-second extension prompts.
• Access to a video generation model, or any image-to-video generator plus FFmpeg for last-frame stitching.
• If your assistant can only chat: describe your product and upload your reference image, then ask the assistant to write the 4 sequential camera-continuous prompts so each segment flows without a hard cut.

Build

1. Read: load the reference image and the 4-stage campaign config file.
2. Decide (continuity check): verify that each extension prompt continues the camera motion and lighting of the previous segment's final frame rather than introducing a new scene cut.
3. Deliver: generate the initial 8-second clip from the reference image, then sequentially call the video extension endpoint 3 times (adding 7 seconds each time) to produce a single continuous 29-second video with native audio. Fallback for models without native extension: extract the final frame of each generated clip with FFmpeg, pass that frame as the starting image for the next prompt, and concatenate the 4 clips with a 0.2-second audio crossfade. Save the final MP4 and prompt manifest locally for review.

Check

• Run only step 1 (the initial 8-second clip) first and inspect the final frame before running the 3 extensions, confirming subject consistency before spending credits on the full 29-second chain.
• Test the fallback with two existing clips: extract the last frame of clip 1, verify its resolution matches the reference image, and stitch without a visible jump cut.
• Quiet run: if the campaign config is missing any of the 4 prompts, the script halts before making any paid video API call.

In my version

• My Python marketing video pipeline, chaining an 8-second initial generation with three 7-second Veo 3.1 extensions from a campaign config file.

Ask first

• What target titles, seniority, location, and compensation floor should the scanner filter for? (default C-level or VP Data/AI roles, remote or local hub)
• What is the minimum fit score out of 10 required before generating a tailored resume? (default 8.0/10)
• May it submit applications automatically? (default never; generate the tailored PDF and application answers for human review and submission)

Needs

• A master markdown file of your career history, verified metrics, and STAR stories (never allow the model to invent facts outside this file).
• A list of target company career portals and a LaTeX or HTML-to-PDF engine.
• If your assistant can only chat: paste your master resume and a job description into chat, and ask for a 1-to-10 fit score, skill gap analysis, and reordered resume bullets using only facts from your master file.

Build

1. Read: scan target ATS portals for new postings since the last run, verify each listing is still live, and deduplicate reposts.
2. Decide (score): evaluate each new role against the master profile on role scope, domain fit, and compensation. Keep roles scoring 8.0 or higher; archive the rest with a one-line reason.
3. Decide (tailor with fact guard): for qualifying roles, select the most relevant experience bullets and stories from the master profile, reorder them for the role, and run a fact-check pass verifying every number and title exists in the master CV.
4. Deliver: compile the tailored PDF resume, draft a short cover letter, and update the local pipeline tracker. Watch email replies to update status when an interview invite or rejection arrives.

Check

• Run the tailor step on a sample job description that asks for a skill not in your master CV: confirm the fact-checker blocks any invented claim and instead notes the gap in the evaluation report.

In my version

• My local career-ops suite with Gemini role evaluation, automated LaTeX/PDF resume generation, strict CV fact verification, and inbox reply watching.