Featured
Signal Vault
A local AI workflow that turned saved Instagram Reels into searchable transcripts, themes, and insights.
- 609
- records
- 500k+
- words
- ~7×
- speedup
Finance student · AI builder
Hi, I'm Logan. I like figuring things out and building better ways to understand them.
I'm in my junior year at NDSU studying Finance with a minor in Banking. My projects usually start with messy information — saved videos, notes, research, or data — and turn it into something clearer and more useful.
- Finance student at NDSU
- Junior year, Banking minor
- Interested in AI, markets, and useful systems
More about Logan
My name is Logan Schiller, a finance student and AI builder focused on systems, data, and business execution. I'm currently studying Finance at North Dakota State University and building projects that use AI to analyze messy information, automate workflows, and surface useful insight.
My recent work includes a 609-Reel AI enrichment project, where I downloaded Instagram Reels, transcribed them, processed the content locally with AI models, and organized the enriched data into a structured Excel dataset.
Selected work
Projects I'm proud of
Origination OS
Flyover
A local-first sourcing system for importing, scoring, contacting, and matching off-market SMB acquisition targets.
- 6-step
- cadence
- 100-pt
- score
- Local
- SQLite app
Flagship case study
Retail Intelligence System
Turned messy Lightspeed POS exports from a seasonal boutique retailer into owner-ready intelligence for inventory, seasonality, staffing, and buying decisions.
Not a generic dashboard: a decision layer for what to clear, what to reorder, when to staff, and what the POS did not directly say.
- ~58%
- frozen cash
- ~87%
- 10am-4pm
- ~40%
- summer sales
- 5.8×
- art vs furniture
Chrome extension
Sports Card Centering Checker
A pre-purchase centering read for collectors evaluating eBay listing photos.
- Score
- per scan
- Signal
- buyer read
- eBay
- screening
How I think
What guides my work
Turning messy inputs into useful systems
I like building simple pipelines that make information easier to find, trust, and act on.
Exploring AI through practical projects
I learn by building. Every project is an experiment in what is actually possible.
Interested in markets, automation, and insight
Those three areas keep me curious and push me to keep improving the systems I build.
Notes & insights
Thoughts & learnings
How I processed 500,000 words of Instagram Reels locally
The full pipeline: download, transcribe, clean, enrich, and export.
Why local AI changed the way I build
Lower cost, more control, better privacy, and more room to experiment.
Lessons from building my first real pipeline
What surprised me about turning scattered content into a useful system.
About
I'm Logan — a finance student who builds practical AI systems for real problems.
I'm a finance student interested in AI, markets, and practical automation. Since ChatGPT first came out, I've used AI to research faster, build small tools, organize messy information, and make better sense of the things I'm working on.
My projects usually start with a real problem: too much saved content, messy data, school work that could be easier, or information that feels useful but hard to act on. I build systems that make those problems clearer.
About me
My journey
-
2022
Self-directed learning
I started exploring AI and new technology on my own — testing tools, following how they worked, and building the habit of learning by experimenting rather than waiting to be taught.
-
2023
Academic foundation
I began my studies at NDSU and started thinking more seriously about real-world problem solving — how business decisions get made, and how a good idea actually becomes useful in practice.
-
2024
Finance in practice
I joined the Student Investment Fund at NDSU and began applying my finance interests in a more serious setting — working with real research, markets, and decisions instead of theory alone.
-
2025
Direction and intent
I became more intentional about where I was headed: pursuing useful certifications, thinking beyond narrow finance roles, and focusing on how companies operate, how decisions get made, and where technology improves them. I also started building small projects with real data.
-
2026
AI systems with real impact
I started building my largest AI projects and applying them to local small-business problems — creating systems that saved real time and supported revenue and operations, not just demos.
- 609
- reels
- 500k+
- words
- ~7×
- faster
Toolbox
Tools I use
Python
Claude
ChatGPT
Ollama
Whisper
Excel
SQL
APIs
Git
Automation
Notes
Notes on what I've built, what I've learned, and how I think through useful AI systems.
This is where I document project breakdowns, practical lessons, and ideas that came out of building with AI. I'm less interested in hot takes and more interested in what actually worked, what became useful, and what helped me understand something better.
Featured notes
Project notes
The pipeline I used to process 609 Reel transcripts
Downloading, transcribing, cleaning, enriching, and turning saved content into something searchable.
Why I started running AI locally
Lower cost, more control, better privacy, and more room to experiment.
What building Signal Vault taught me
Lessons from turning scattered content into a system I could actually use.
Things I've learned
Practical takeaways
Better structure beats more information.
Most useful insight comes from organizing what already exists, not endlessly collecting more.
Small workflows become real leverage.
The biggest wins often come from simple repeatable systems, not complicated software.
Clear questions create better outputs.
When I define the problem well, the tools become much more useful.
Local tools make experimentation easier.
Running things privately and cheaply makes it easier to test more ideas.
Documentation improves the work.
Writing down the process usually reveals what matters and what still needs work.
Useful systems start with a messy input.
Saved content, notes, transcripts, and raw data become valuable when the structure is good.
Note library
Browse by topic
Contact
If you're working on an interesting problem, I'd be glad to connect.
I'm especially interested in AI, research, automation, markets, and practical systems that make messy information easier to use.
Email[email protected]
LocationMinnesota, USA
Featured project
Signal Vault
A local-first AI pipeline that turned saved Instagram Reels into transcripts, structured fields, scoring signals, and a searchable Excel dataset.
Overview
What it is, and why it matters.
What it is
A local AI pipeline that converted saved Instagram Reels into searchable transcripts, structured fields, scoring signals, and an Excel dataset.
Why it matters
Saved short-form video holds useful ideas, but it is hard to search, compare, prioritize, or turn into action without structure.
What changed — the archive became a private decision-support layer for learning, building, filtering ideas, and spotting repeated themes.
In short
Goal, build, and result.
Make the archive searchable.
Turn saved reels into text and fields that could be reviewed without rewatching everything.
Run the pipeline locally.
Download, transcribe 500k+ words, enrich with Llama 3.1, score, and export every record.
A usable dataset.
609 structured records, 100% enrichment completion, $0 marginal API cost, and ~7× faster inference.
Problem
Saved videos are useful, but not searchable.
Instagram Reels can contain tool ideas, business angles, workflows, and market signals. Saving a reel keeps the video, but it does not capture the idea in a way that can be searched, compared, filtered, or reused.
Passive archive
609 short videos, scattered context, and no reliable way to compare ideas across creators or topics.
Queryable system
609 transcripts, 500k+ words, and enriched records organized into a dataset that can be filtered and scored.
System
The pipeline
Four stages move a saved reel from raw video to a structured, scored record. Select a stage to see what it does.
Python and yt-dlp pulled down saved Instagram Reel videos using cookie-authenticated access and per-record state tracking.
Optimization
Two decisions made the pipeline practical.
Swapped Qwen 3 for Llama 3.1
Why
Qwen 3 kept entering reasoning-heavy extraction behavior at roughly 150 seconds per record. This task needed clean field extraction, not extended reasoning.
Result
About 7× faster inference and more consistent JSON output with Llama 3.1 8B.
Local-first, not an API
Why
609 LLM calls on a paid API can add up and would move private saved-content data off-machine.
Result
$0 marginal cost, fully private processing, and repeatable enrichment on demand.
Intelligence layer
The schema made the archive useful.
{
"creator": "...",
"topic": "local AI workflow",
"content_type": "AI_TOOL",
"concepts": ["agent workflows", "local inference"],
"tools": ["Ollama", "Whisper"],
"opportunities": ["automation service for SMBs"],
"future_signals": ["on-device models"],
"skills": ["Python", "prompt engineering"],
"actionability": 8,
"leverage": 9,
"novelty": 7
}Every reel became comparable by creator, topic, concept, opportunity, future signal, skill, and score.
The schema turned a pile of transcripts into rows that could be searched, filtered, ranked, and reviewed.
Actionability, leverage, and novelty made the dataset useful for deciding what was worth acting on.
Dataset use
The dataset became a decision-support layer.
Searchability
Turned video content into searchable text.
Structure
Converted messy reels into consistent rows and columns.
Prioritization
Scored ideas by actionability, leverage, and novelty.
Pattern detection
Surfaced recurring themes across hundreds of saves.
Decision support
Helped decide what to learn, build, ignore, or act on.
Results
Measured output
Conclusion
The dataset turned saved content into a usable decision system.
Signal Vault was not just a transcription project. It created a structured dataset that can be searched, filtered, scored, and reviewed to decide what to learn, build, ignore, or act on.
Featured system
Flyover
A local-first origination OS for off-market SMB deal sourcing. Flyover turns a static target list into a working dispatch system: which companies to contact, why they rank highly, what touch is due next, and which buyers might fit if an owner replies.
Overview
An operating system for the part of deal sourcing that usually breaks.
Flyover was built around a simple operating problem: target lists are easy to create, but hard to work consistently. The project turns a list of potential SMB acquisition targets into a dispatch system with priority, cadence, replies, and buyer fit.
What it is
A local web app for managing off-market SMB sourcing from target import through buyer matching.
Why it matters
The value is knowing who deserves the next touch, what needs to happen today, and how to act when an owner replies.
What changed — a spreadsheet became a workflow system with one clear next action per active company.
In short
Goal, build, and result.
Make sourcing operational.
Move off-market SMB sourcing out of scattered spreadsheets, notes, and memory.
Create a local workflow app.
Import targets, score conviction, schedule outreach, track replies, and match buyer mandates.
A dispatch-first system.
Every active company has a next action, every score has a reason, and every reply can move toward buyer fit.
CSV importConviction scoring6-step cadenceDispatch board
Reply handlingBuyer matchingLocal-first architecturePytest workflow rules
Problem
Sourcing does not usually fail from a lack of targets. It fails from unclear follow-up.
Static list
Unclear priority, manual follow-up, forgotten next steps, no reply workflow, and no clean buyer handoff.
Dispatch system
Ranked target universe, scheduled cadence, sent/skipped/replied history, buyer matching, and explainable score.
System flow
From target universe to buyer match.
Import
Bring in a CSV target universe with flexible header matching.
Score
Rank companies using transparent conviction factors.
Dispatch
Start cadences and work from the due-today board.
Reply
Cancel pending touches when an owner responds.
Match
Compare the company against buyer mandates with rationale.
Dispatch board
The dispatch board is the product.
Flyover is designed around daily action. Instead of asking the user to browse a database, it shows what must go out today, ordered by conviction score.
Dispatch · Today
Highest conviction sends first
Fictional demo records
ScoreCompanyStepChannelDueAction
94Prairie Fire Protection1/6lettertodaySent
91Clearwater Well Drilling2/6emailoverdueSkip
88Iowa Falls Machine & Tool3/6calltodaySent
84Heartland Foodservice Supply4/6emailtodaySent
Conviction scoring
The score is transparent by design.
The scoring layer does not try to be magic. It creates a priority system the operator can inspect, argue with, and adjust.
{
"owner_tenure": "long-held businesses score higher",
"owner_age": "succession risk signal",
"succession_signal": "retirement or no-heir evidence",
"industry_durability": "essential local services rank higher",
"revenue_fit": "lower-middle-market target bands",
"geography": "Upper Midwest focus",
"digital_weakness": "less visible businesses surfaced"
}Conviction Score0-100
Owner tenure20
Owner age band14
Succession signal15
Industry durability15
Revenue fit15
Geography10
Digital weakness5
Cadence engine
One invariant drives the workflow: every active company has a next touch.
Letter→Email→Call→Email→Call→Letter→Re-engage
Creates the first pending touch.
Schedules the next touch immediately.
Records the skip, then continues the sequence.
Cancels pending touches and moves to human judgment.
Kills the company record.
Moves dormant with a 180-day re-engage touch.
Buyer matching
A reply should become a buyer conversation quickly.
When an owner responds, the next question is not where the buyer spreadsheet went. It is which buyer mandate fits this company and why.
Company profile
Prairie Fire Protection
fire protection · SD · 3_10m revenue band
Matched buyer
Granite Pass Partners
Fit rationale- Industry match
- Revenue band in mandate
- In territory
- Fee agreement in place
Example uses fictional demo data from the seed file.
Technical architecture
Built local-first on purpose.
CSV Import→SQLite→FastAPI routes→Python modules→Jinja UI→Pytest invariants
SQLite
A one-person sourcing operation needs a reliable local file, not unnecessary cloud infrastructure.
Jinja UI
The product is form-driven and workflow-heavy, so server-rendered templates keep state simple.
FastAPI
Clear routes and form actions for queue, universe, company, buyer, and import workflows.
Pytest
The repo includes tests for scoring determinism, pending-touch invariants, reply cancellation, queue ordering, and match rationale.
What I built
The execution spans backend logic, workflow rules, and operator-facing screens.
Importer
Flexible CSV header matching and duplicate skipping by company name.
Scoring engine
Explainable 100-point conviction model with factor-level breakdowns.
Sequencer
Cadence state machine with next-touch enforcement.
Dispatch board
Daily action surface ordered by conviction.
Company page
Score breakdown, touch history, reply logging, and buyer matches.
Buyer book
Mandate storage with fit rationale for warm replies.
Design decisions
Product judgment mattered as much as code.
Dispatch-first, not database-first
Why
A company list only matters if it creates action.
Transparent scoring, not black box scoring
Why
A sourcer needs to trust and adjust the priority model.
Local-first, not SaaS-first
Why
The first version is for a single operator and should not require cloud setup.
State machine over reminders
Why
The workflow needs rules: sent schedules next, reply cancels pending, skip is recorded.
What this demonstrates
A finance workflow turned into software.
Reflection
Flyover is strong because it is narrow.
It does not try to become a full private-equity platform. It focuses on the part of sourcing that breaks first: consistent follow-up. The project turns a scattered target list into an operating rhythm, where the next action is always visible and the logic behind each priority is inspectable.
Featured project
Sports Card Centering Checker
A Chrome extension that helps collectors evaluate card centering directly from eBay listing photos before buying. It turns a raw marketplace image into a visual centering workspace, then translates the alignment into a readable centering score, buyer signal, and risk flag.
Overview
A focused tool for one specific buying decision.
Collectors often judge centering from marketplace photos by eye, but small centering issues can affect perceived condition and grading potential. This extension adds a clearer pre-purchase read directly where the buyer is already evaluating the card.
What it is
A Chrome extension that takes eBay listing images of cards into a workspace where it gives an accurate centering score.
Why it matters
It gives buyers looking for mint-condition cards that could grade high more confidence in one of the hardest parts of grading.
What changed — instead of guessing from a listing photo, the buyer can align the card edge and printed border, then get a readable centering score, buyer signal, and risk flag.
In short
Goal, build, and result.
Help collectors judge centering before buying.
Make one important condition factor easier to evaluate from imperfect eBay listing photos.
Add an edge-to-border alignment layer.
Inject a browser-based centering workspace directly onto listing images.
A focused pre-purchase screening tool.
A raw listing photo becomes a clearer centering read that helps buyers make better decisions.
Decision layer
The alignment becomes a scan-based buyer read.
Generated from the current listing image and edge-to-border alignment.
- Horizontal Centering
- Measured from alignment
- Vertical Centering
- Measured from alignment
Problem
Listing photos are useful, but centering is hard to judge by eye.
eBay photos are often enough to see the card, but not enough to confidently compare how the printed design sits inside the physical edge. The tool makes that visual judgment more structured before a collector commits to a purchase.
Raw listing photo
Visual information is present, but the buyer still has to estimate centering without measurement structure.
Pre-purchase centering read
The same photo becomes easier to interpret with an alignment workspace and buyer-facing output.
System
Listing photo → alignment workspace → buyer-facing read.
Capture listing image
Start from the card photo being evaluated.
Align edge and border
Align the physical card edge and printed design border.
Calculate centering balance
Compare horizontal and vertical edge gaps.
Return scan-based read
Translate the measurement into a clearer buyer-facing centering output.
Measurement workspace
The real overlay keeps the visual judgment on the listing image.
Card edge guide
The outer guide maps to the physical card edge.
Design border guide
The inner guide maps to the printed border inside the card.
Horizontal and vertical balance
Centerlines and labels show how the spacing is distributed.
Manual alignment layer
The buyer adjusts the guides because marketplace photos vary in crop, angle, cases, glare, and lighting.
Technical build
Built as a local Chrome extension, not a hosted app.
Manifest
MV3 extension configured for eBay listing pages and image-hosting domains.
Content script
Runs on eBay pages, identifies the main visible image, and injects the trigger button.
Canvas overlay
Draws the card image, alignment guides, handles, labels, and centering stats in one controlled workspace.
Interaction
Supports guide handles, background pan, scroll zoom, fit-to-screen, and reset behavior.
Measurement logic
Centering comes from comparing edge gaps.
The tool compares the distance between the outer card edge and the inner printed border on each side. Left and right gaps become horizontal centering. Top and bottom gaps become vertical centering. The output becomes a practical centering read from the available listing image.
{
"outer": "physical card edge",
"inner": "printed design border",
"horizontal": "left gap / right gap",
"vertical": "top gap / bottom gap",
"output": "centering score + buyer signal"
}The measurement starts with user alignment because listing photos vary in angle, crop, lighting, glare, cases, and stands.
The centering read gives a clearer signal, but it does not grade corners, edges, surface, authenticity, or final value.
Design decisions
The useful part was keeping the tool narrow.
Manual alignment was intentional
Why
Listing photos vary too much in angle, lighting, holders, glare, and crop quality for a simple auto-detect promise.
Impact
The user stays in control and can align the guides to the visible evidence in the listing photo.
Overlay directly on eBay
Why
The tool is most useful at the moment a collector is evaluating a listing.
Impact
No separate upload flow is required before getting a centering read.
Canvas-based interface
Why
Canvas made it easier to draw the image, guide lines, dashed borders, handles, labels, and crosshairs together.
Impact
The measurement layer feels like one controlled workspace instead of separate floating DOM pieces.
What this demonstrates
A small tool with a real workflow behind it.
Extension architecture
Packaged a narrow utility as a browser-native workflow.
DOM injection
Found and enhanced the listing image without requiring a separate app.
Canvas drawing
Rendered image, guides, handles, labels, and lines inside one measurement layer.
Coordinate transforms
Kept measurement points aligned through zoom, pan, and screen-to-image conversion.
Product focus
Solved one collector workflow without overstating what the tool can do.
Reflection
This project works because it solves one narrow problem cleanly.
Sports Card Centering Checker does not try to become a full grading platform. It takes one judgment collectors already make by eye and gives them a structured way to make that judgment more consistently from the listing image in front of them.
Case study · Research + workflow design
Making AI useful for people who don't have time to figure it out
Almost everyone has met AI. Almost no one has been shown what to do with it. This is a look at the gap between access and usefulness — and the one skill that closes it: turning a messy, repeated task into a workflow someone can actually reuse.
Overview
What I studied, and why it matters.
What I studied
Why people have a near universal access to AI but still struggle to make it useful. and what actually changes when they do.
Why it matters
The people who could benefit most are small teams, local businesses, nonprofits, students, solo operators who all have the least time to experiment their way to a system.
What I used
Current sources I trust on this: Pew, the U.S. Census Business Trends survey, McKinsey and HBR, Microsoft's Work Trend Index, MIT, the SBA, NFIB, and a 2026 field experiment from INSEAD.
The core insight
The next useful AI skill isn't prompting. It's workflow translation: taking a repeated task and turning it into a reliable, reusable system.
The thesis in one line — the gap isn't access to AI. It's the missing step between a tool and a workflow, and that step can be designed.
In short
The question, the approach, the answer.
Why does access not equal usefulness?
People try a chatbot once, get a generic answer, and quietly go back to doing the task by hand.
Read the real evidence, then design.
Synthesize current adoption and barrier research, then turn the pattern into a framework and a buildable concept.
Design the workflow, not the prompt.
Usefulness comes from translating a repeated task into inputs, instructions, a review step, and a reusable output.
The core problem
Access ≠ usefulness.
Access is basically solved: the models are cheap or free and built into the tools people already use. But usefulness leaks out at every step between “I have AI” and “AI helps me get this done.” The funnel doesn't convert. It drains.
StartAccess to an AI tool
Leak 1Unclear use case
Leak 2Weak first attempt
Leak 3No repeatable workflow
EndAbandoned
This isn't only a beginner problem. MIT's NANDA initiative found that roughly 95% of enterprise generative-AI pilots produced no measurable impact — and concluded the cause was approach, not model quality. If well-funded companies leak at the workflow step, a two-person shop has no chance without help.
Who this affects
Fluent, still stuck
Heavy users, but often using AI to finish work instead of to learn it and have no model of what good looks like.
Admin overload, no IT
NFIB found only 24% of small employers use AI and 42% aren't familiar with it. No time to experiment, real worry about accuracy.
One person, a chatbot
High adoption, low impact usually one staffer drafting with AI while everyone else stays buried in manual work.
No slack to redesign
Microsoft found 80% of workers lack the time or energy for the work which is the very condition that prevents fixing it.
Every hour counts
They are the whole company. Every hour spent figuring out AI is an hour not earning, so the bar for “worth it” is brutal.
The framework
Six things that decide whether AI is actually useful.
“Is this AI any good?” is the wrong question. The better question is whether the work around the model is set up well. These six dimensions are how I judge that and each comes with the question to ask yourself.
01
Problem clarity
A model can't hit a target you can't name. Vague task in, vague answer out.
Ask: can I state the task in one sentence?
02
Input & context quality
Most “AI failures” are missing-context failures. The model can't know what you didn't give it.
Ask: does it have the facts this needs?
03
Human judgment
The person decides what matters, what's true, and what's safe to send. The model proposes; the human disposes.
Ask: who is checking this, and against what?
04
Output usefulness
Useful output is in the format you'll actually use, not a wall of text you have to rework.
Ask: can I use this as-is, or near it?
05
Iteration
The first draft is a draft. Anthropic's 2026 fluency analysis found iteration was the strongest predictor of getting real value.
Ask: did I refine, or accept the first answer?
06
Real-world application
It only counts if it produces a decision or an artifact you act on and if you can run it again next week.
Ask: did this change what I do, repeatably?
From prompt to value
The value isn't in the model. It's in the chain around it.
The same six dimensions show up as a process. Select a stage to see its real job and notice how little of the value lives in the AI draft itself.
It starts with a real question or task, not a tool. “Write me something” goes nowhere. “Reply to this refund request in our voice, under 120 words” gives the model a target.
The point — five of the six steps are human work. The AI draft is the cheap part. Skip the human steps and you get fast, confident, unusable output.
What the research shows
Five findings, and what each one means for design.
I didn't want a pile of statistics. I wanted the few findings that actually change how you'd build something. Expand each for why it matters and what it implies.
A 2026 INSEAD field experiment gave 515 startups identical AI tools, training, and credits. The only group that pulled ahead was the one told how to reorganize their work around AI — they found 44% more use cases and generated roughly 1.9× the revenue of peers who used the same tools just to go faster.
For design: don't teach tools. Help people find and re-shape the repeated task.
INSEAD working paper, 2026Anthropic's 2026 fluency analysis found that 85.7% of substantive conversations involved refining a previous answer — and that iteration predicted every other good behavior. The people who get value treat AI as a back-and-forth, not a vending machine.
For design: build the loop in. A review-and-refine step means users don't have to know to do it themselves.
Anthropic AI Fluency Index, 2026The gains are real when AI is embedded in a defined task — a 14% average productivity lift in customer support, and 34% for the least-experienced workers. But HBR and McKinsey note that verification and rework quietly eat those savings when there's no checkpoint.
For design: a defined check isn't friction. It's what makes the speed bankable.
NBER (Brynjolfsson et al.); HBRAcross small-business surveys, the top barriers are accuracy, data privacy, no clear use case, and the learning curve — not price. The SBA found 82% of firms with fewer than five employees say AI simply “isn't applicable” to them, which reads as a use-case gap, not a true mismatch.
For design: you can't price past distrust. Show what good looks like, keep a human in the loop, and say when not to use AI.
SBA Advocacy (2025); SMB surveysAdoption tends to be wide but shallow: U.S. Census data shows most AI-using firms apply it in three or fewer functions. Tools don't adopt themselves — usually it's one motivated person, and the workflow leaves when they do.
For design: make the workflow itself the owner. A reusable artifact carries the instructions so it doesn't depend on one person.
U.S. Census BTOS, 2025–26The difference
Basic AI use vs. a useful AI workflow.
Same tool, same model, completely different result. The difference is entirely in the work around the prompt.
A blank box and a vague question
- Ask a broad question
- Accept the first answer
- No context or evidence provided
- Nothing reusable next time
- Output is hard to verify or trust
A defined task with a recipe
- Define the specific task
- Provide the context it needs
- Check assumptions and facts
- Iterate to a better version
- Save it as a reusable artifact
- Output supports a real decision
One worked example
A salon answering the same DMs all day. Basic use: type “reply to this” and hope. Designed workflow: a saved template with the services, prices, hours, and tone, plus “never invent a price — leave a [CHECK].” Paste the message, get a draft, glance at it, send. Five minutes of writing becomes thirty seconds of checking, and it works the same way tomorrow.
Try it
Score a workflow you actually run.
Pick a repeated task where you already use AI and rate it on the six dimensions. Nothing is pre-filled — this is your read of your own workflow. The summary updates as you go.
Clear problem
Can you state the task in one sentence?
Good context
Does the model get the facts it needs each time?
Verifiable output
Can you quickly tell whether it's right?
Human judgment
Is there a defined moment where a person checks it?
Reusable process
Can you run it again next week in under a minute?
Practical result
Does it end in a decision or artifact you act on?
Rate each part of a workflow you actually run — the summary updates as you go.
Reusable prompts
Templates, not blank chats.
The blank box is the enemy. These are starting points for the most common repeated tasks — copy one and fill in the brackets.
Turn messy notes into next steps
Here are my raw notes from [meeting / call / brainstorm]: [paste notes]. Turn them into: (1) decisions made, (2) action items with an owner and a due date, (3) open questions, (4) risks to watch. If something is unclear or missing an owner, list it under "needs clarification" instead of guessing.
Create a customer reply template
You're helping me answer customer messages for [business type]. Our tone is [warm / professional / casual]. Key facts: [hours, prices, policies]. Here's a reply I was proud of, as a style example: [paste]. Now draft a reply to this message: [paste]. Keep it under [X] words. Never invent prices, dates, or policies — if you don't have a fact, leave a [CHECK] placeholder.
Turn a repeated task into a workflow
I do this task repeatedly: [describe the task]. Help me turn it into a reusable workflow. Ask me for: (1) what a good result looks like, (2) what inputs it needs each time, (3) any constraints or things to never do. Then give me: a reusable instruction block, a short review checklist, and one example. Keep it simple enough that a busy non-expert could run it.
Check this AI output before I use it
Review this AI-generated [email / summary / post] before I use it: [paste]. Check for: (1) factual claims that might be wrong or unverifiable, (2) anything confidential or sensitive, (3) tone mismatches for [audience], (4) anything overstated or invented. List issues as a short checklist with fixes. Don't rewrite it unless I ask.
What I'd build
A concept: an AI Workflow Starter Kit.
This part is a prototype direction, not something I've shipped. If the gap is translation, then the thing worth building isn't another AI tool — it's the missing layer between a tool and a workflow. A small, opinionated kit that does the translation for nontechnical people.
Start from the task
Intake
A five-minute form: “what do you do over and over that you wish took less time?” It diagnoses the repeated task instead of asking people to imagine use cases.
Cards
One workflow card per common job — customer reply, meeting to tasks, feedback synthesis — each carrying the instructions, an example, and a review checklist, pre-filled.
Build in trust
Guardrails
A plain-language “when not to use AI” note on every card — legal advice, sensitive data, high-stakes facts. Naming the no-go zones is what earns trust for everything else.
Library
A saved set that grows over time, so the workflow belongs to the team and not to one motivated person who might leave.
Encouragingly, this is already a fundable need: NTEN's free “AI for Nonprofits” cohort, supported by Anthropic, exists specifically to walk stretched, nontechnical teams through launching one real AI project — proof the translation layer is real, not just an idea.
What this demonstrates
What I hope this shows about how I work.
Systems thinking
I see AI adoption as a workflow-design problem, not a tool problem.
Workflow design
I turn a vague idea into a framework and a buildable concept.
Research synthesis
I find current, credible evidence and notice when two numbers measure different things.
Clear communication
I can make a messy topic scannable without dumbing it down.
Practical judgment
Small first wins, human review, reusable templates, and naming the limits.
Takeaway
The interesting problem isn't AI's capability. It's the distribution of usefulness.
The models are already good enough. The gap is that the people who'd benefit most have the least time to figure it out, and almost no one is doing the unglamorous translation work that would help them. That's the reframe I keep coming back to: stop teaching people to prompt, and start helping them turn a repeated task into a system they can actually use.
Sources
Where the numbers come from.
Awareness vs. use of AI: 95% of US adults have heard of AI; about a third have used a chatbot, and only a third of those found it very helpful.
The rigorous adoption floor: ~18% of US firms used AI in a business function, and most use it in three or fewer functions.
The GenAI Divide: ~95% of enterprise generative-AI pilots showed no measurable P&L impact; the cause was approach, not model quality.
Field experiment, 515 startups: the group that reorganized work around AI earned ~1.9× the revenue of peers with identical tools. (Working paper, not yet peer-reviewed.)
AI Fluency Index: 85.7% of substantive conversations involved iteration, the strongest predictor of effective use.
Small-firm reality: 24% of small employers use AI, 42% aren't familiar with it, and 82% of sub-five-employee firms say AI “isn't applicable.”
Work Trend Index: 80% of workers say they lack the time or energy for the work; interruptions hit roughly every two minutes.
Brynjolfsson et al.: 14% average support-productivity gain (34% for novices); HBR on how rework erodes savings without a review step.
Figures are reported as the sources state them. The INSEAD result is a 2026 working paper with partly self-reported revenue, so I've flagged it rather than leaning on it as settled fact.