aaron

Quick win (5 minutes): Create one Google Sheet and add a single row: today’s date, one headline you saw this morning, source, URL, and tag it with the topic you care about. That tiny habit is your pipeline starter.

The problem: Signals are everywhere — news, social, forums — and without a simple system you either miss early wins or drown in irrelevant noise.

Why this matters: A repeatable, low-cost trend pipeline converts scattered signals into prioritized experiments you can run. Speed and discipline beat perfect coverage.

My experience — short lesson: I’ve built lightweight systems for teams that don’t have engineers. The single lever that creates value is consistency: one spreadsheet as the truth plus a weekly AI synthesis that turns rows into prioritized experiments.

1. What you’ll need
   • Google account (Sheets, Alerts, Trends)
   • RSS reader or an email folder for clippings
   • Access to an AI assistant (ChatGPT or similar)
   • Optional: Zapier/Make for automating new rows into the Sheet
2. How to set up (step-by-step)
   1. Create one Sheet with columns: date, source, headline/snippet, URL, tag, sentiment, independent-signals (count), priority (1–5), notes.
   2. Pick 3–4 focused topics. For each, choose 5–8 high-value sources (news, one subreddit, one Twitter/X list, one newsletter).
   3. Set Google Alerts + add RSS feeds. If you don’t automate, forward interesting items to one inbox and paste into the Sheet daily.
3. Daily collection & triage
   1. Skim alerts, add rows to the Sheet, tag topic and sentiment.
   2. When an item appears across different sources, increment independent-signals. Require ≥3 signals before flagging a trend candidate.
   3. After two weeks, prune sources that rarely contribute unique signals.
4. Weekly synthesis & action
   1. Export the week’s rows and run the AI prompt below to surface 3–5 trends.
   2. Score candidates with Reach×Impact/Effort (1–5). Pick the top 1–2 for simple experiments (landing page, 5 outreach emails, quick interviews).
   3. Run one measurable test per week and log results back into the Sheet.

Copy-paste AI prompt (use weekly)

“You are an analyst. Here are short snippets (date, source, headline/snippet, URL). Identify up to 5 emerging trends across these items. For each trend provide: title (5 words max), 2–3 supporting signals from the snippets, a confidence score (0–100), business implications (3 bullets), one recommended next experiment (one sentence). Present output as a numbered list.”

Metrics to track

• Signals captured per week
• Trend candidates identified per month
• % of candidates with ≥3 independent signals
• Experiments launched per month
• Conversion: trend → validated opportunity → revenue

Common mistakes & fixes

• Chasing single-source noise — fix: require 3 independent signals before escalation.
• Too many sources — fix: keep the top 6 that give unique value after two weeks.
• No action — fix: limit to one clear experiment per week with one metric.

1-week action plan

1. Day 1: Create the Sheet, define 3 topics, set 5–8 sources and Google Alerts.
2. Days 2–5: Collect items daily into the Sheet (5–10 rows/day ideal).
3. Day 6: Run the AI prompt on the week’s rows; pick top trend using Reach×Impact/Effort.
4. Day 7: Design and launch one small experiment; measure one metric and record outcome.

Your move.

Quick win (5 minutes): Ask an AI to generate 5 thumbnail concepts and 3 color palettes for one niche. Pick the best and upload a single mockup to a POD listing to test demand.

Nice focus in your thread title on “reliably” selling — that’s the right KPI. Many people chase looks instead of predictable returns.

Problem: designers treat POD like art instead of a repeatable product funnel. That means one-hit wonders and wasted ad spend.

Why it matters: a repeatable process gets you consistent sales, predictable margins, and the ability to scale. With AI you can cut research and design time from days to hours.

Experience in a sentence: I’ve built POD lines where 30% of SKUs drive 80% of revenue — and those winners came from systematic niche testing, fast iteration, and tight listing optimization.

1. What you’ll need
   • AI image tool (text-to-image or vector generator)
   • Basic mockup tool (POD platform or simple editor)
   • Spreadsheet to track tests
   • Marketplace (Etsy, Redbubble, Merch, etc.)
2. Step-by-step process
   1. Pick one niche (e.g., “funny gardening mugs”): spend 10 minutes validating by searching marketplace top sellers.
   2. Generate 10 design concepts with AI. Use the prompt below. Expect 20–60% to be usable after quick edits.
   3. Create 3 polished mockups per design (shirt, poster, mug). Keep images high-res and on transparent backgrounds for POD.
   4. Publish 5 listings with strong, keyworded titles and 3 targeted tags each.
   5. Run paid boost or organic promotion for 7–14 days, track traffic and conversions, then scale winners.

Copy-paste AI prompt (use as-is):

Create 10 variations of a minimalist botanical line-art design for a 12×12 inch print and a standard mug. High-contrast black lines on a transparent background, vector-style, limited 2-color palette options. Provide 3 color palette suggestions and 3 short title ideas for each design suitable for POD marketplaces.

Metrics to track

• Listing views
• Click-through rate (CTR) from search or ads
• Conversion rate (views → sales)
• Profit per sale and return on ad spend (ROAS)
• Repeat purchase / customer reviews

Common mistakes & fixes

• Poor mockups → fix: use lifestyle images and clear close-ups.
• Low-res assets → fix: export vector or 300 DPI PNGs.
• No keyword testing → fix: iterate titles/tags weekly and watch CTR.
• Too broad niche → fix: narrow to specific subgroups (hobby + demographic).

1-week action plan

1. Day 1: Pick 1 niche, run quick marketplace scan (30 min).
2. Day 2: Use AI prompt to generate 10 concepts; shortlist 5 (1 hour).
3. Day 3: Create 3 mockups per shortlisted design (1–2 hours).
4. Day 4: Write optimized titles/descriptions/tags for 5 listings (1 hour).
5. Day 5–7: Launch listings, promote with a small budget or social post, track metrics daily.

Your move.

— Aaron

Good point — the trigger + human review combo is the backbone. I’ll add the KPI focus and a clear, non-technical rollout you can run this week.

The gap

Many teams automate follow-ups but measure only sends and opens. That hides whether you’re actually moving deals forward or building trust with useful resources.

Why this matters

If your follow-ups increase replies and resource clicks, you shorten sales cycles and reduce wasted outreach. If they don’t, you’re just sending noise and training people to ignore you.

My experience in one line

Start with a tight loop: AI drafts → human reviews → automation sends → track 3 KPIs → iterate weekly. That takes the guesswork out and protects relationships.

What you’ll need

• Recipient list + last message or meeting note (paste the last exchange).
• Objective per recipient: help / nudge / book call.
• Resource library: 3–6 vetted links or attachments by topic.
• An AI tool (ChatGPT or equivalent) and your email tool with template/trigger capability.
• A reviewer for high-value prospects (1 person, 10–15 minutes per email).

Step-by-step rollout (what to do, how to do it, what to expect)

1. Choose 20 priority prospects. Export names, companies, last message.
2. For each, select one objective and 2–3 relevant resources from your library.
3. Run the AI prompt below to generate 2 subject lines and 2 email variations.
4. Human-review: check facts, add one personal detail, approve or edit (5–10 mins).
5. Set automation: trigger = 3 days no reply; embargo manual send for top 20% value.
6. Expect: first batch sent Day 3, measurable replies within 48 hrs of send.

Copy-paste AI prompt (primary)

Act as a helpful business follow-up writer. Using this context: [paste last email or meeting note], recipient name: [Name], role: [Role], company: [Company], objective: [help / nudge / book call]. Produce: 2 subject lines; Version A (concise, 3–4 sentences) and Version B (resource-first, 6–8 sentences). For each version include 2–3 suggested resources with one-line value statements and a single clear CTA (reply or schedule). Keep tone friendly, professional, under 150 words. Flag missing personalization fields.

Variants

• Short variant: ask for a one-line reply instead of a call.
• Resource-first variant: open with the most relevant link and why it matters.

Metrics to track (start here)

• Reply rate (primary): target +10–20% lift vs current baseline.
• Resource click rate: target ≥25% of opens click a link.
• Meetings booked: measure closed-loop from reply → booked call.

Common mistakes & fixes

• Too many resources — fix: 2 max, ranked by likely relevance.
• Auto-send without review — fix: manual embargo for top-tier prospects.
• No CTA or confusing CTA — fix: one clear action (reply or schedule).

7-day action plan

1. Day 1: Collect 20 prospects + 10 resources by topic.
2. Day 2: Run prompt on 5 samples and review outputs.
3. Day 3: Finalize templates and subject lines in your email tool.
4. Day 4: Configure automation triggers and manual embargo rules.
5. Day 5: Send first batch (with review). Monitor replies/clicks.
6. Day 6: Tweak prompts/resources based on early data.
7. Day 7: Roll to next 50 prospects or adjust cadence.

Clear, measurable steps — minimal tech required, maximum control. Your move.

Hook: Want a convert-ready product video in a day using Runway and a couple of AI tools? Do this with minimal kit and no editing anxiety.

The problem: Most people overthink production—too many shots, long scripts, bad audio. That kills attention and ROI.

Why this matters: Attention is short. A clear 20–30s video that shows the benefit, demonstrates usage, and has captions will outperform a pretty but vague 60s film. Fast execution wins.

What I’ve learned: Ship a simple, benefit-first video, then iterate. The first objective is measurable engagement—views, clicks, and add-to-cart lifts—not cinematography awards.

What you’ll need (simple):

• Smartphone or one product clip
• Runway account (or similar editor)
• Short script: 20–45 words
• 1 music bed, logo, product photos

Step-by-step (do this now):

1. Decide one message: problem + single benefit + CTA.
2. Write a 25–30s script: 3s hook, 15–20s demo/benefit, 4–6s CTA.
3. Record three clips: close-up product (3–5s), product in use (4–7s), smiling user or logo end-card (3–5s).
4. Open Runway: import clips, trim to beat, remove background if required, assemble timeline.
5. Add captions: auto-generate and tighten line length to 3–6 words per line.
6. Voice: use Runway TTS or a quick recorded VO. Mix voice + light music at -12 to -18dB under voice.
7. Export 9:16 for social and 16:9 for web. Preview on phone and desktop.

What to expect: A lean 20–30s asset that conveys the benefit, works muted, and drives clicks. First version will be rough—that’s fine.

Metrics to track:

• View-through rate (VTR) at 3s and 15s
• Click-through rate (CTR) from video to product
• Conversion rate on landing page (post-click)
• Cost per purchase if running ads

Common mistakes & fixes:

• Too much info: Cut to the single benefit. Edit ruthlessly.
• Bad audio: Use TTS or re-record voice in a quiet room.
• No captions: Always add; most watch muted.
• Long intro: Start with the pain or hook in first 2–3s.

Copy-paste AI prompt (use to generate scripts & shot lists):

“Write three 25–30 second product video scripts and 3-shot storyboards for a wireless charging dock called FloatCharge. Tone: clear, benefit-first, slightly friendly. Each script: 1) a 3-second hook, 2) a 15–20 second demo/benefit, 3) a 4–6 second CTA. For each script list exact caption lines (max 6 words each) and camera framing per shot (close-up, medium, wide).”

1-week action plan (exact):

1. Day 1: Run the prompt, pick best script, record three clips.
2. Day 2: Edit in Runway, add captions, apply TTS or VO.
3. Day 3: Export 9:16 and 16:9, preview on devices, fix timing.
4. Days 4–7: Run A/B test: two hooks or two CTAs. Measure VTR and CTR.

Small experiments to run: Test two hooks and two CTAs. Keep everything else identical; pick winner by CTR and VTR.

Your move.

Nice addition — the triple-lock (FAQ retrieval, slot-filling, confidence thresholds) is exactly the guardrail you need. I’ll add a tighter, KPI-focused plan so you get measurable results fast and obvious next steps.

Problem: automation that saves time but makes avoidable mistakes — or nothing measurable changes because you didn’t track it.

Why it matters: reduce first-response time, protect CSAT, and recover billable hours. If you automate the right intents you can cut 30–60 minutes/day on a busy week and hit 60% automation on safe intents within 3 weeks.

Lesson: retrieval-only + slot checks + per-intent thresholds create predictable outcomes. Don’t guess policy — force the bot to source answers and collect required data first.

Do / Don’t checklist

• Do: Build a one-page FAQ and map 5–10 canned replies with placeholders.
• Do: Require slot completion before policy replies.
• Do: Start in suggest-mode and log every decision.
• Don’t: Auto-send refund or legal language.
• Don’t: Let the model invent policy — use retrieval-only prompts.

Step-by-step setup (what you’ll need, how to do it, what to expect)

1. What you need: one-page FAQ, 5–20 canned replies, 10–50 example messages per intent, a helpdesk or chatbot with prompt hooks and confidence scores.
2. Configure: implement triage prompt (classify intent + list slots), then retrieval-only answer linked to your FAQ, then a tone-polish step.
3. Thresholds: OrderStatus auto-send ≥92%, ReturnRequest auto-send ≥90% with slots, RefundRequest & TechnicalIssue remain suggest-mode initially.
4. Test: run 20 real messages, measure mismatches, tweak phrasing and slot lists.
5. Go live: 2-week suggest-mode trial, then promote 3–5 safe intents to auto-send.

Copy-paste AI prompt (use exactly as written)

You are my customer support triage assistant. Classify the message into one of: OrderStatus, ReturnRequest, RefundRequest, TechnicalIssue, BusinessHours, Other. List which required details (slots) are present and which are missing. If confidence ≥ 90% and all required slots are present, return the matching canned reply with placeholders filled. If confidence is 60–89% or any slot is missing, draft a single friendly clarifying question to collect missing details. If confidence < 60% or the message shows urgent/angry tone, recommend escalation. Return: intent, confidence (0–100), found slots, missing slots, and either a filled reply or one clarifying question.

Metrics to track

• First Response Time (target < 1 hour)
• Automation Rate (% messages auto-sent)
• Escalation Rate (target < 15%)
• CSAT / thumbs-up rate
• Time saved per week (minutes)

Common mistakes & fixes

• Over-automating sensitive cases — Fix: block refunds/legal from auto-send.
• Poor slot collection — Fix: require slot-first clarifier before any policy reply.
• Model invents policy — Fix: use retrieval-only prompt tied to your FAQ.

Worked example — OrderStatus flow (quick ROI)

1. Customer: “Where’s my order #123?”
2. Triage: Intent=OrderStatus, Confidence=95%, slots present: order number, email.
3. Retrieve: pull shipping status from FAQ/DB and fill reply template.
4. Auto-send: “Your order #123 shipped on 20 Nov via Carrier X; expected delivery 24–26 Nov. Track here: [Link].”
5. Result: OrderStatus often 40–60% of volume — promoting it to auto-send saves 15–30 minutes/day.

7-day action plan (exact steps)

1. Day 1: Create one-page FAQ and 5–10 canned replies.
2. Day 2: Define slots and escalation triggers.
3. Day 3: Install triage + retrieval-only + tone prompts in tool.
4. Day 4: Test with 20 past messages; log results.
5. Day 5: Go live in suggest-mode; review every suggestion.
6. Day 6: Promote top 3 safe intents to auto-send.
7. Day 7: Measure metrics, add 10 new examples per intent, adjust thresholds.

Your move.

Agreed: your time-boxed workflow and stop rules are exactly the discipline most teams skip. I’ll layer on two upgrades: a message library that makes AI copy sharper, and guardrails that protect lead quality so wins actually move revenue.

The problem

Headline tests often “win” on clicks but quietly hurt quality or fail to replicate. Teams celebrate too early, then watch downstream metrics slide.

Why it matters

Reliable lifts require believable claims with proof and a simple safety net. Done right, you get compounding 5–20% gains without burning traffic or brand trust.

What experience taught me

AI is best at speed and breadth. The compounding wins come from: 1) feeding AI real customer language, 2) using a tight Claim–Proof–Action structure, and 3) enforcing guardrails (lead quality, traffic balance, time in market).

Do this next — step-by-step

1. Set KPIs and guardrails (15 minutes)
   • Primary KPI: conversion rate to your core action (e.g., qualified lead submission).
   • Guardrails: bounce rate, CTA click-through, and lead quality (e.g., MQL rate or booked-call rate). Define “acceptable” bands (e.g., no worse than -5% vs control).
   • Stop rules: minimum 7–14 days, 100+ conversions per variant, 95% confidence before calling a winner.
2. Build a Message Bank (30–45 minutes)
   • Collect 10–20 snippets from customer reviews, call notes, sales emails, or support tickets.
   • Tag each snippet as pain, outcome, objection, or proof (metrics, logos, quotes).
   • Pick your top 3 outcomes and 3 proofs. These power your copy and your tests.
3. Generate structured variants with AI (20 minutes)
   • Use a Claim–Proof–Action pattern. Don’t let AI ramble — force constraints.
   • Ask for three directions: clarity-first, urgency-first, and social-proof-first. Keep the rest of the page identical.
4. Optional pre-screen (cheap signal, 1–2 hours)
   • Run a small ad or on-site poll to gauge first-click interest on headlines only. Spend a small, fixed budget.
   • Advance only the top 1–2 variants to the A/B test.
5. Launch the A/B test (30 minutes)
   • Even traffic split. One variable only (e.g., headline + subhead).
   • QA: confirm pixels fire once, goals track, and that 50/50 traffic is actually ~50/50 (sample ratio mismatch check).
6. Monitor sanely, decide once
   • Check daily but don’t call it early. Respect the stop rules and guardrails.
   • If variant wins and guardrails hold, ship it as the new control. If quality drops, discard and test the next direction.

Copy-paste AI prompt — Message Bank to variants

“Act as a senior conversion copywriter. Using the inputs below, produce three landing page variants using a Claim–Proof–Action structure. Constraints per variant: 8–10 word headline, 15–22 word subhead stating the main outcome, two 8–12 word bullets tying features to outcomes, a 2–3 word CTA, and one short social-proof sentence. Create three angles: Clarity-first, Urgency-first, Social-proof-first. After the copy, write a one-sentence test hypothesis and list two guardrail metrics with acceptable ranges. Inputs: Persona = [describe], Primary pain = [describe], Desired outcome = [describe], Key proof (metrics/quote/logo) = [paste], Offer = [describe], Top objections = [list]. Tone: confident, plain English, non-technical. Audience: over-40 decision-makers.”

Copy-paste AI prompt — A/B test queue

“You are a CRO strategist. Given: Baseline conversion = [x%], Weekly unique visitors = [#, by source], Primary KPI = [define], Guardrails = [define], Current headline = [paste], Current CTA = [paste]. Propose 5 single-variable test hypotheses with: expected impact (H/M/L), rationale (1 line), required minimum conversions per variant (use simple 100–200 min rule if unsure), and an implementation note. Prioritize using ICE (Impact x Confidence x Ease) and return the ICE score.”

What to expect

• Headline-only tests typically move 5–20% on conversion rate when the claim is clearer and proof is visible above the fold.
• Not every “win” sticks — guardrails prevent quality erosion. Expect 1 in 3 to 1 in 4 tests to be a keeper. That’s normal.
• Documented learnings turn single wins into a playbook you can reuse across pages and segments.

Metrics to track

• Primary: conversion rate to your core action.
• Guardrails: bounce rate, CTA CTR, lead quality (e.g., MQL rate or booked-call rate).
• Directional: time on page, scroll depth, revenue per visitor (if applicable).
• Health check: traffic split balance (if 50/50 deviates by more than ~3 points, investigate tagging or allocation).

Common mistakes and quick fixes

• Mistake: Calling a result mid-week. Fix: Always run in full-week increments (7 or 14 days) to capture weekday patterns.
• Mistake: “Winning” on clicks but losing on quality. Fix: Require guardrails to be within your preset band before declaring a winner.
• Mistake: Sample ratio mismatch (uneven splits). Fix: Check allocation and ensure each user is bucketed once; verify one pixel fire per event.
• Mistake: Low traffic, underpowered tests. Fix: Test bigger changes (claim + subhead), aggregate pages with similar intent, or extend duration.

1-week action plan

1. Day 1: Lock KPI + guardrails; pull 7–14 days of baseline by source.
2. Day 2: Build the Message Bank from real customer language; pick top 3 outcomes and proofs.
3. Day 3: Run the Message Bank prompt; select two variants (clarity vs proof).
4. Day 4: Implement A/B (even split), QA tracking, confirm traffic balance.
5. Day 5: Launch; commit to stop rules (min 7–14 days, 100+ conversions/variant).
6. Day 6: Monitor guardrails only; do not call the test.
7. Day 7: Decide, ship the winner if quality is stable, and document the learning. Queue the next test (CTA or testimonial placement).

Keep your workflow tight: message-proof first, single-variable tests, and guardrails that protect outcomes. Do that, and AI becomes a force multiplier instead of noise.

Your move.

Hook: Want cold emails that get replies — not polite ignores? Send fewer words, one measurable outcome, one concrete ask.

The core problem: most cold emails ramble, list features, and ask for vague time. Recipients don’t see a reason to reply.

Why this matters: reply rate is the fastest lever to increase qualified meetings and predictable pipeline. Small improvements scale: a 5–10% lift in reply rate doubles your meetings if you keep send volume steady.

Short lesson from the field: I switched a campaign from 6-paragraph pitches to a 3-sentence, outcome-led format. Reply rate doubled and meeting quality improved — because the ask was simple and the value measurable.

Do / Do not — quick checklist

• Do: keep it 3 sentences, state one measurable outcome, offer two specific slots.
• Do: personalize one sentence with public, recent context.
• Do not: cram benefits or jargon into the first message.
• Do not: use private or creepy personal details.

What you’ll need

• List of 20–50 prospects: name, title, company, one-line public context.
• Simple tracking (spreadsheet/CRM): send date, reply, meeting booked.
• ChatGPT (or similar) to create 3 subject lines + 2–3 body variants per prospect.

Step-by-step (do this every batch)

1. Pick one target outcome (e.g., reduce churn 5% / increase MRR by X%).
2. For each prospect capture one-line context (recent article, product, metric you can see publicly).
3. Use this 3-sentence template: 1) one-line connection, 2) one-line outcome/value, 3) one-line CTA with two time options.
4. Run the AI prompt below to generate subjects and 2–3 body variants; pick the most human-sounding.
5. Send first 20 emails, follow up at day 3 and day 7 with one-line nudges and a single time option.

AI prompt (copy-paste):

“Write three cold-email variants (each 3 sentences) to request a 15-minute exploratory call about improving client retention by 5% for a mid-market SaaS VP of Customer Success. Include one sentence showing a personal connection (use: [insert personalized context here]), one clear value statement tied to a measurable outcome, and one simple call-to-action proposing two specific time slots (example: Tue 11am or Thu 2pm). Also provide 3 short subject lines.”

Worked example — copy, tweak, send

Subject: Quick 15 mins to reduce churn by 5%?

Hi [Name], I liked your recent post about onboarding improvements — good practical points. We help mid-market SaaS teams remove the top two onboarding drop-offs and lift retention ~5% within 90 days. Any chance for 15 minutes — Tue 11am or Thu 2pm?

Follow-up (day 3): Still open to a quick 15-minute chat to review two easy retention wins? Tue 11am or Thu 2pm?

Metrics to track

• Reply rate (replies / sends) — primary KPI.
• Meeting conversion (meetings / replies).
• Pipeline value from meetings (estimate).

Common mistakes & fixes

• Too many benefits — fix: pick one measurable outcome and remove the rest.
• Personalization feels creepy — fix: limit to one public detail and cite the source (article, product release).
• Weak CTA — fix: offer two concrete times or one simple next step.

1-week action plan

1. Day 1: Build 20–50 prospect list + one-line context.
2. Day 2: Generate variants with the prompt; choose best subject + body.
3. Day 3: Send first 20 emails.
4. Day 6 and 10: Send short follow-ups to non-responders.
5. Day 11: Review reply and meeting rates; iterate subject/body based on top performers.

Your move.

5‑minute win: Paste your last 30 merchant names into an AI chat and ask it to build a “vendor map” (merchant → default category + follow‑up question). Save that map. It will auto‑classify 80% of future transactions and shrink your monthly review to minutes.

The real problem: you’re losing deductions and time because expenses aren’t consistently categorized, mixed‑use items aren’t split, and evidence isn’t captured. That’s money left on the table and higher audit friction.

Why this matters: a tight, AI‑assisted workflow increases legitimate deductions, reduces prep time by half, and gives your tax preparer a clean package. That’s less back‑and‑forth and more dollars retained.

What you’ll need

• CSV exports (date, description, amount, merchant).
• Scans of key receipts (PDF/JPG) and any mileage notes.
• Home office details (dedicated sq ft and total home sq ft).
• A spreadsheet plus any AI chat that accepts pasted text.

How to systemize your deduction hunt

1. Define categories you’ll reuse all year. Advertising/marketing, software, supplies, professional fees, education, utilities, phone/internet, meals, travel, home office, mileage/vehicle, other. Keep them identical to what your filing software uses for smooth handoff.
2. Build two maps: Vendor Map + Exceptions Map. Vendor Map = recurring merchants and default categories with a one‑line “business‑purpose question.” Exceptions Map = obvious NON‑DEDUCTIBLE patterns (e.g., transfers, credit card payments, groceries). These two maps prevent 90% rework.
3. First AI pass with your maps. Run the CSV through AI to categorize using the Vendor Map, apply Exceptions Map, and mark ambiguous items as REVIEW with a reason.
4. Apply mixed‑use splits once, reuse monthly. Phone/internet/software bundles: set a conservative business‑use % (e.g., 60%) and store it. AI applies it going forward and notes your rationale.
5. Vehicle + home office: choose a draft path. Pick mileage or actual expenses for the car and stick to one. For home office, have AI compute both a simplified and an actual‑style draft so you can pick one to discuss with your preparer.
6. Evidence memos for big/ambiguous items. For anything that looks personal or high‑value, paste the receipt text for an AI‑generated memo: who/why/date. Save memo + receipt together.
7. Variance + duplicate sweep. Ask AI to compare month‑to‑month totals, flag spikes, duplicates, reimbursements, and transfers. Clean before totaling.
8. Export a tidy package. Category totals, REVIEW list with your notes, mixed‑use % summary, vehicle/home‑office choice, and a folder with CSV + receipts + memos. That’s your audit‑ready bundle.

Robust AI prompt — maps + categorize + clean

“I’m a freelancer reviewing expenses. Here are CSV rows with headers (date, description, amount, merchant). 1) Build a VENDOR MAP: recurring merchant, common aliases, default category (marketing, software, supplies, professional fees, education, utilities, phone/internet, meals, travel, home office, mileage/vehicle, other), and a short ‘business‑purpose question’ if meals/travel/ambiguous. 2) Build an EXCEPTIONS MAP: patterns to tag as NON‑DEDUCTIBLE (transfers, credit card payments, refunds, owner draws, reimbursements) and the detection rule you used. 3) Using both maps, categorize each transaction, mark uncertain items as REVIEW with the reason, and tag NON‑DEDUCTIBLE where appropriate. 4) Suggest conservative business‑use % for mixed‑use items and show both gross and deductible amounts. 5) Output: a) totals by category (deductible only), b) a REVIEW list with the exact follow‑up question I must answer, c) a NON‑DEDUCTIBLE list (transfers/refunds/etc.) so I can exclude before filing. Do not give tax advice; just categorize, flag, and total.”

Prompt — receipts to audit memo

“Extract from this receipt text: date, vendor, amount, item(s), payment method. Create a 1‑line business purpose memo in plain English. If personal or unclear, label REVIEW and ask me the missing detail. Output as: RECEIPT MEMO: [one line].”

Prompt — mileage and home office drafts

“Here are my trips (date, start/end, purpose, miles). Total monthly and year‑to‑date business miles. Flag any trip missing purpose as REVIEW. I also have X sq ft dedicated office and Y sq ft total home. Provide a simple draft comparison of a simplified home‑office estimate vs. an actual‑style approach in plain English with the inputs I gave. Label outputs as drafts to discuss with my tax preparer.”

What to expect

• 80–90% auto‑categorized after your first maps; 10–20 minutes to clear the REVIEW list monthly.
• Cleaner totals and fewer follow‑ups from your preparer.
• Consistent mixed‑use logic and less risk of double counting.

Metrics that matter

• Automation rate: % auto‑categorized (target ≥85%).
• Time spent per month (target: under 60 minutes).
• New deductions surfaced ($) vs. last period.
• Receipt coverage: % of high‑value/ambiguous items with a memo (target 100%).
• Reclassification rate after preparer review (track to improve your maps).

Mistakes and fast fixes

• Mixing car methods. Fix: declare “mileage” or “actual” in your prompt; tag the other bucket as NON‑DEDUCTIBLE.
• Ignoring mixed‑use items. Fix: set a conservative % once; store rationale; apply automatically.
• Counting transfers/refunds. Fix: formalize an Exceptions Map; auto‑exclude before totals.
• Thin evidence. Fix: generate a one‑line memo for every big/ambiguous item; save with the receipt.
• Unreconciled income. Fix: ask AI to compare income deposits vs. your client/platform list and flag gaps for review.

7‑day action plan (hands‑on, minimal tech)

1. Day 1 (30 min): Export last 2 months of CSVs. Paste top 30 merchants into the maps prompt; save Vendor + Exceptions Maps.
2. Day 2 (30–45 min): Run the categorize prompt on 100–200 rows. Mark REVIEW items; add one‑line notes.
3. Day 3 (20 min): Set mixed‑use % for phone/internet/software; note your rationale.
4. Day 4 (20 min): Paste key receipts into the memo prompt; attach memos to files.
5. Day 5 (20 min): Run the variance/duplicate sweep (re‑run categorize prompt asking for anomalies). Clean.
6. Day 6 (20 min): Run the mileage + home office draft prompt; choose drafts to discuss with your preparer.
7. Day 7 (20 min): Export totals by category, REVIEW list, and package receipts/memos into one folder. Set a monthly reminder.

Experience/lesson: with clients, the combination of a reusable Vendor Map, an Exceptions Map, and short evidence memos is the difference between clever AI output and an audit‑ready package. Build once, reuse all year.

Your move.