Jeff Bullas

Quick win: You don’t need to be an SEO expert to make your Etsy and Shopify listings sell better — you need a clear process and one good AI prompt to get you started.

Why this matters: Etsy and Shopify look similar but use keywords differently. Treat each listing like two tasks: one for Etsy (titles, tags, first lines) and one for Shopify (meta, URL, alt text). Use AI to speed brainstorming, not to replace your product facts or voice.

• What you’ll need:
  • Product facts (materials, size, color, use, packaging)
  • 3–5 seed keywords or customer phrases
  • 1 clear product photo
  • Access to your shop analytics or note of best sellers

1. Gather your product facts and 3 seed keywords — think like a buyer (why they buy, for whom, occasions).
2. Ask AI for related search phrases, 4–6 title options, 10 tag suggestions and a short product description (first 1–2 lines + bullet specs + call-to-action).
3. Edit AI output to match facts and your tone. Put highest-value keywords in the first 60–80 characters of titles and meta titles.
4. Platform polish: Etsy — fill tags and use exact-phrase variations; Shopify — set meta title/meta description, URL slug and image alt text with top keywords.
5. Launch & measure: publish, monitor impressions/clicks/conversions, tweak every 2–4 weeks.

Worked example (quick): handmade linen tea towel. Seed keywords: “linen tea towel”, “kitchen towel”, “housewarming gift”.

• Title options (pick one): “Linen Tea Towel – Soft Natural Kitchen Towel – Housewarming Gift”, “Handmade Linen Kitchen Towel | Eco-Friendly Dish Cloth”
• First lines for Etsy: “Handmade linen tea towel — absorbent, quick-dry and soft. Perfect for everyday kitchen use or as a thoughtful housewarming gift.”
• Shopify meta: Meta title: “Linen Tea Towel — Soft Eco Kitchen Towel”; Meta description: “Handmade linen tea towel, absorbent and quick-dry. Ideal as a housewarming gift. Free domestic shipping over $50.” Alt text: “linen tea towel natural color folded on wooden table”

Common mistakes & fixes:

• Do not copy-paste the same copy across platforms — adapt for Etsy tags and Shopify meta.
• Do not stuff keywords unnaturally — make titles readable first, searchable second.
• If impressions are high but conversions low — improve photos, first 1–2 lines and price clarity.

Practical AI prompt (copy-paste):

“You are a friendly product listing assistant. Given this product: [PRODUCT NAME]. Facts: [MATERIALS], [SIZE], [COLOR], [USE/OCCASION], target buyer: [WHO]. Seed keywords: [KEYWORD1], [KEYWORD2], [KEYWORD3].

Please provide: 5 short title options (keywords up front), 12 Etsy tag suggestions, a 2-line Etsy opening description that answers ‘what it is’ and ‘why it matters’, a 4-bullet spec list, a short call-to-action, and a Shopify meta title (under 60 characters), meta description (under 160 characters), and one image alt text. Keep tone warm and clear.”

1. Action plan: Pick one product, run the prompt, paste results into your listing and edit facts/voice.
2. Test: Check analytics after 2 weeks and tweak top-performing keywords.

Want me to run this for one of your listings? Tell me the product name and 2–3 seed keywords and I’ll give three practical edits to try next.

Spot on about the 5-minute validation. That single rule keeps AI fast and keeps you in charge. I’ll add a pro move: build a reusable differentiation blueprint and run a tight two-prompt loop. That’s how you get consistent, classroom-ready lessons in minutes without losing quality.

Why this works: AI is great at structure. Your expertise provides the guardrails: objective, student tiers, time, and materials. Combine both and you’ll get clear tracks, printable tasks, and data you can act on next lesson.

What you’ll need

• One clear objective and the standard text (paste it in).
• Three ability tiers (Remedial / On-level / Extension) from recent work or a quick pre-check.
• Lesson time (e.g., 45 minutes), materials you actually have, and any non-negotiables (no devices, limited printing).
• 5–10 minutes to scan for accuracy, age-appropriateness, and timing fit.

Step-by-step: the differentiation blueprint loop

1. Snapshot first: List who’s in each tier based on last exit ticket or a 3-question pre-check.
2. Generate the lesson: Use the blueprint prompt below to get 3 tracks, timing cues, and printables.
3. Add management details: Ask for grouping, transitions, and what you do while each group works.
4. Validate in 5: Check one worked example per track, language level, and that tasks fit your time.
5. Teach and collect data: Use the exit ticket; note which students were in each track.
6. Refine: Feed the results into the refinement prompt to tighten next lesson.

Copy‑paste AI prompt: Differentiation Blueprint (reusable)

“You are an instructional designer for mixed-ability classrooms. Create a [LESSON_LENGTH]-minute Grade [GRADE] lesson on [TOPIC], aligned to this standard: [PASTE STANDARD TEXT]. Produce three tracks: Remedial (clear visuals + guided practice), On-level (scaffolded mixed problems), Extension (challenge or mini-project with real-world context). Include: 1) student-friendly objective, 2) 5-minute hook, 3) 25-minute activities split by track with timing cues, 4) 10-minute whole-class plenary, 5) a 5-question exit ticket with answers (2 core, 2 application, 1 error analysis), 6) materials list limited to what I have: [LIST MATERIALS], 7) teacher moves for running three groups, 8) one worked example per track, 9) language at Grade [GRADE] reading level. Constraints: no tech beyond [ALLOWED TOOLS], printable tasks fit on 2 pages total, avoid jargon. Output in clean sections I can print.”

Insider tricks that save time

• Constraints as guardrails: Tell the AI exactly what you can print, what tools students have, and your room layout. It prevents unusable plans.
• Reading-level control: Ask for sentence length and vocabulary caps to keep directions clear for lower readers.
• Attention resets: Insert 60–90 second micro-breaks every 12–15 minutes in the prompt; it boosts on-task behavior.
• Accommodations baked in: Request 2–3 universal supports (sentence frames, manipulatives, chunking) per track.

Copy‑paste AI prompt: Roster → Tiers (optional, fast)

“Given this list of students with pre-check scores and brief notes: [PASTE SCORES/NOTES], group them into three tiers (Remedial, On-level, Extension). For each tier, list 2 strengths, 2 needs, and the right track from the next lesson. Flag any students who may need additional scaffolds (e.g., visuals, sentence frames). Keep privacy and tone respectful.”

Copy‑paste AI prompt: Refinement loop (feed it your results)

“Here are exit-ticket results and quick observations from today’s lesson on [TOPIC]: [PASTE RESULTS + NOTES]. Revise tomorrow’s 45-minute plan using the same three tracks. Tighten misconceptions you see (name them), adjust problem difficulty by tier, replace any slow activities with faster alternatives, and keep printing to 2 pages. Provide: updated activities with timing, 5 new exit-ticket questions with answers, and one 3-minute reteach mini-lesson for the most common error. Keep directions at Grade [GRADE] level.”

Mini example (what to expect)

• Remedial: Visual model, 8 guided problems, think-aloud script, sentence frames, manipulatives list.
• On-level: Short worked example, 12 mixed problems, pair-check protocol, timing cues.
• Extension: Real-world challenge or mini-project, rubric bullets, 2 stretch questions.
• Plenary: Whole-class error analysis from the exit ticket.

Mistakes and quick fixes

• Too much to print: Cap printables to 2 pages and ask the AI to compress.
• Over-differentiating: Keep three tracks only; avoid micro-variants that add management load.
• Vague timing: Demand minute-by-minute cues and cut any activity that exceeds your block.
• Reading overload: Tell the AI to limit directions to short sentences and include icons or bullets.
• Extension = more of the same: Ask for transfer tasks or real-world applications, not just harder numbers.

Quick action plan

1. Today: Run the Differentiation Blueprint prompt with your next objective; validate in 5–10 minutes.
2. Tomorrow: Teach it and collect the exit ticket; note who was in each track.
3. Next day: Use the Refinement loop prompt with your results to tighten the follow-up lesson.

Bottom line: Keep it simple. One blueprint, three tracks, five minutes of validation, and a tight feedback loop. That’s how you get reliable differentiation without burning your planning time.

Cheering you on — try the blueprint today and iterate tomorrow.

Jeff

Quick refinement first

Nice checklist — one small tweak: don’t just “store it where AI can read it.” Make the SOW and weekly inputs structured (table, CSV or clearly labeled bullets). AI works far better with predictable fields: deliverable name, baseline hours, acceptance criteria, requester, date. That reduces false positives and speeds review.

Why this approach works

Catch new asks early, convert them into clear change orders, and keep margins intact. Start simple, automate the parts that remove grunt work, and keep a short human review in the loop.

What you’ll need

• Canonical SOW in a simple table: deliverable | baseline hours | acceptance criteria.
• Weekly inputs: meeting bullets (date, requester, ask), timesheet totals by deliverable.
• A single folder or spreadsheet to collect inputs.
• A lightweight AI assistant (chat or batch analyzer) that can read your structured files.
• Change-order template and a client message template.

Step-by-step setup

1. Export the SOW to a table or CSV. Fill baseline hours per deliverable.
2. Standardize meeting notes: one bullet = date | requester | short ask | related deliverable.
3. Pick two flags to start: (A) New deliverable name not in SOW; (B) New hours > 10% of baseline or +8 hours.
4. Each week, paste the SOW table and that week’s bullets into the AI tool. Run the prompt below.
5. AI returns flagged items and a draft change-order. Project lead reviews within 24 hours, adjusts costs, and sends the client message.
6. Only update the SOW after client approval and signed change order.

Example flag & draft change order

• Flag: Add mobile onboarding flow — not in SOW; estimated 16 hours (baseline 0). Change-order draft: “Client requested a new mobile onboarding flow. This is outside current SOW and requires 16 hours of additional work. Estimated cost $1,600. Please confirm approval to proceed.”

Common mistakes & fixes

• Noisy inputs — fix: enforce the date|requester|ask|deliverable format.
• Unclear baselines — fix: require baseline hours before starting a package.
• Blind trust in AI — fix: always human-review flagged items within 24 hours.

Copy‑paste AI prompt (use as-is)

Compare the following SOW table and weekly meeting bullets. For each meeting bullet that is not in the SOW or increases estimated hours by more than 10% (or +8 hours), list: 1) short description of the change; 2) new deliverable or scope expansion; 3) baseline hours and new estimated hours; 4) percent change; 5) recommended time and cost impact; 6) a concise change-order draft (2–4 sentences) and a short client message with accept/decline options. Output structured bullets.

1‑week action plan

1. Day 1: Convert SOW to a table and set the two flag rules.
2. Day 2: Create meeting-note template and change-order template.
3. Day 3: Run the prompt on last week’s notes; record flags.
4. Day 4: Review flags, send 1st change-order if needed.
5. Day 5: Measure flags, approvals, time-to-decision; tweak threshold if false positives >30%.

Closing reminder

Start small, automate the comparison, keep the human check. Preventing one big scope surprise pays for the whole system.

Short answer: Yes. AI can quickly spot duplicate tasks, cluster similar work, and help you cut redundancy so your team spends time on outcomes, not repetition.

Why this matters: Over time task lists bloat. Meetings, emails, and checklists create a jungle of repeat actions. A simple AI-assisted review surfaces repeats, consolidates effort, and frees hours each week.

What you’ll need

• A consolidated task list (CSV, spreadsheet, or exported from your tool).
• Access to an AI tool that can process text (chat or API).
• Basic spreadsheet skills or a project-management view to apply changes.

Step-by-step: how to do it

1. Gather: Export tasks from tools (Trello, Asana, email, calendar) into one spreadsheet with columns: Task, Owner, Frequency, Context.
2. Normalize: Remove trivial differences (lowercase, trim dates) and add short tags like “reporting,” “follow-up.”
3. Ask AI to cluster: Use the prompt below to group similar tasks and mark duplicates.
4. Review results: Human-in-the-loop—confirm groupings, merge tasks, assign a single owner and cadence.
5. Implement: Update your workflow, set recurring tasks once, and remove redundant steps or tools.
6. Automate detection: Add a weekly script or AI check to flag new duplicates as your list grows.

Copy-paste AI prompt (use as-is)

Here is a list of tasks, each on its own line. Group them into sets of duplicates or near-duplicates and explain in one sentence why they are the same. Suggest one consolidated task label and recommended owner/recurrence. Tasks:
[PASTE YOUR TASK LIST HERE]

Prompt variants

• Short variant: “Cluster these tasks by similarity and propose one consolidated task for each cluster with owner and frequency.”
• Spreadsheet variant: “Given columns Task, Owner, Frequency, Context, identify duplicate tasks and add a column ‘ConsolidationID’ to group duplicates. Explain rules used.”

Example

Input tasks: “Send weekly sales report”, “Prepare weekly sales dashboard”, “Email weekly sales numbers to execs” → AI groups them and suggests: “Weekly sales report (Owner: Sales Ops, Recurrence: weekly).”

Common mistakes & quick fixes

• Relying on AI 100% — Always review suggested merges before deleting tasks.
• Poor data quality — Clean the spreadsheet first (fix typos, standardize phrasing).
• Ignoring context — Keep a context column so similar-sounding tasks with different goals aren’t merged wrongly.

7-day action plan

1. Day 1: Export and clean your task list.
2. Day 2: Run the AI clustering prompt and review results.
3. Day 3–4: Consolidate tasks, assign owners and set recurrence.
4. Day 5: Remove or archive redundant steps and update SOPs.
5. Day 6–7: Schedule a weekly quick AI check or automation to catch new duplicates.

Final reminder: Start small. Consolidating a few repetitive tasks creates immediate time savings and builds momentum for bigger workflow redesigns. Use AI to surface patterns — you decide what changes.

Nice — I like the focus on a short, repeatable setup. That 30–90 minute workflow is exactly the right energy: fast, practical, and human-led. Here are a few additions to make it even more foolproof when you hand things to AI.

What you’ll need (short checklist)

• One-page voice guide (5–10 tone words + 3–5 example sentences).
• Voice bank of 10 favourite lines saved in one folder.
• Channel list with length goals (email, social, ads, support).
• Templates: 3 per channel (draft, subject/headline, reply script).
• A simple review rule: human checks first 20, then spot-checks.

Do / Don’t checklist

• Do give the AI a sample sentence to match each time.
• Do ask for 3 variations and pick the best.
• Don’t over-correct—teach with one clear edit at a time.
• Don’t assume a single prompt will be perfect forever—refresh examples monthly.

Step-by-step (30–90 minutes)

1. Create the one-page voice guide and paste 3–5 live example sentences from your best content.
2. Make a voice bank of 10 lines and label them by channel/use.
3. Build three templates per channel. Note the single goal for each template (click, reply, resolve).
4. Use the AI with the prompt below to generate 3 versions. Pick one, give one small tweak, and save the final as a template.
5. Repeat for one channel this week. Add top outputs to the voice bank and schedule a 15-minute monthly refresh.

Worked example — GreenLeaf Coffee (quick)

1. Voice guide: warm, helpful, slightly cheeky; examples: “We make great coffee easy.” “Quick brew tips you’ll actually use.”
2. Channel: Instagram caption, 20–30 words, goal = click to shop.
3. Ask AI for 3 captions, choose the friendliest, tweak one word to be simpler, save as template.

Common mistakes & fixes

• Mistake: Vague prompt. Fix: always include a sample sentence and goal.
• Mistake: Over-editing outputs. Fix: prefer small, specific feedback like “make it 20% warmer.”
• Mistake: No versioning. Fix: save originals + final so you can retrain if tone drifts.

Copy-paste AI prompt (use as your baseline)

Match this brand voice and output three variations:
Tone words: warm, confident, plain English.
Reference sentence: “We make great coffee easy.”
Channel: Instagram caption, 25 words max, goal: drive click to shop.
Constraints: no slang, use one short question, include a clear CTA.
Output: label variations A, B, C. After each, suggest one small tweak to make it warmer.

Action plan for today: 1) build the one-page guide (15–20 min), 2) create voice bank (15 min), 3) run the prompt for one channel and save best output (20–40 min). Small wins stack fast — start with one channel and scale.

Remember: the secret is examples + tiny corrections. Get the AI to mimic 10 great lines and it will keep your brand sounding the same everywhere.

Quick win (try in 5 minutes): Pick one recent email, run the AI prompt below to generate five subject lines and three matching preheaders, choose one pair, and swap them into your next send.

Why this matters: Sender name + subject + preheader (the “Header Stack”) gets the open. The first 400px of the email gets the click. Fix both and you’ll see CTOR and revenue move — not just opens.

What you’ll need

• Access to your ESP with A/B testing and mobile preview
• A simple AI chat tool for copy options
• Brand tokens: sender name format, 2 color hexes (primary + contrast), 1 hero image
• Metrics to track: unique clicks/delivered, CTOR, unsub rate

Step-by-step (do this)

1. Choose one control email (your last send that performed “average”).
2. Run the AI prompt below to get subject + preheader pairs and CTA options.
3. Pick a mobile-first one-column template: headline, hero image, one primary CTA (button), one tiny secondary link.
4. Make the CTA obvious: 44px min height, 16–18px type, contrasting color, 1–2px border for dark mode.
5. Test ONE variable: either A/B subjects (same body) or A/B CTA color/copy (same subject). Send to 10–20% sample.
6. Measure by unique clicks/delivered and CTOR. Push winner to the rest and log results.

Example (ready to copy)

• Sender: Alex at Brand
• Subject: Save 30 mins today — quick fix (38 chars)
• Preheader: One simple routine to cut time on admin (46 chars)
• Headline: Stop wasting time — try this 3-step fix
• CTA: Try it now (button color: #FF6A00; border: 2px #FFFFFF)

Common mistakes & fixes

• Too many CTAs — fix: one primary button, one small text link.
• Testing multiple variables — fix: test only one at a time.
• Invisible CTA in dark mode — fix: add border and test in dark preview.
• Gmail clipping — fix: keep HTML & images under ~90KB.

1-week action plan

1. Day 1: Run the AI prompt below. Pick 4 subject/preheader pairs.
2. Day 2: Build two variants (same body, two subjects).
3. Day 3: Send A/B to 15% sample. Wait 24–48 hours.
4. Day 4: Push winner to the rest. Log in your Wins Library.
5. Day 5: Use an AI Template Critique prompt on the winning email and apply 2 fixes.
6. Day 6–7: Test CTA copy or button color. Review CTOR and revenue.

Copy-paste AI prompt (use in your chat tool)

Act as a senior email strategist. Create 8 subject lines (28–44 characters) and 6 matching preheaders (35–70 characters) for an email about: [describe your offer and audience]. Requirements: lead with a clear benefit or specific detail (number or timeframe), friendly tone, no spammy words, at most one tasteful emoji. Group as subject + matching preheader pairs. Also provide 4 short CTA button texts (1–3 words) and 2 sender name formats that increase recognition.

What to expect: Small, repeatable lifts — look for a 5–10% CTOR bump within two cycles. If you don’t see improvement, rewrite the headline to include a number, timeframe or outcome and rerun the test.

Reminder: Speed matters. Generate options, test fast, measure by clicks, and roll winners forward into your Wins Library.

Spot on about the 5‑minute runway check and the Monte Carlo caution. Timing is the real killer in cash flow, and many chat AIs won’t simulate thousands of runs unless you feed them a sheet. Let’s use AI where it shines: building the structure fast, so your spreadsheet does the heavy lifting.

Try this in under 5 minutes

• Ask AI to generate a ready-to-paste 13-week cash flow CSV with built-in scenario toggles. Paste it into your spreadsheet, drop in a few numbers, and you’ll see your “breach week” and top cash levers immediately.

Copy-paste prompt (use as-is)

“Create a 13-week cash flow model as CSV text. Columns: Week, Opening Cash, Cash In (AR Current), Cash In (AR 30), Cash In (AR 60), Cash In (Other), Total Cash In, Payroll, Rent, Subscriptions, Taxes, Debt Service, Vendor Payments, CAPEX, Other Outflows, Total Cash Out, Net Cash Flow, Closing Cash. Include a Parameters section at the top with: Min Cash Threshold, AR Collection Assumptions (percent collected by bucket and week), Vendor Payment Lag (weeks), Stress Factors (Revenue % change, AR Days +/-, One-off Cash Outflow in Week N). In the CSV, add formulas (A1-style) so Closing Cash = Opening Cash + Total In – Total Out, and Opening Cash rolls forward each week. Add a small Summary block that calculates: Runway in weeks until Closing Cash < Min Cash Threshold, Top 3 drivers by variance vs Base, and a short list of 4 actions with estimated cash impact and timing (e.g., tighten AR by 5 days, defer CAPEX, negotiate vendor terms, early-payment discount).”

What you’ll need

• Last month’s ending cash and your minimum cash threshold (e.g., two payrolls + rent + tax set-aside).
• 12–24 months of monthly cash in/out or, for speed, just the next 4–8 weeks of known inflows/outflows.
• Your AR aging (Current/30/60/90) and payroll/vendor schedules.

Build it — step by step

1. Generate the CSV with the prompt above and paste into your spreadsheet (most tools accept paste-to-grid).
2. Set Parameters: Opening Cash, Min Cash Threshold, and simple AR collection rules (e.g., 70% current, 20% in 30 days, 10% in 60 days).
3. Enter known events: payroll dates, rent, taxes, debt service, big invoices due, and any one-offs.
4. Switch on two scenarios: Pessimistic (–20% Cash In, AR +7 days) and Optimistic (+10% Cash In, AR –5 days). The summary should show breach week by scenario.
5. Ask AI to translate any missing logic into formulas if your sheet needs help (e.g., rolling Opening Cash, stress toggles).

Premium insight: model timing, not just totals

• Cash improves fastest when you shift when money moves. Add lags: AR days, vendor terms, implementation delays for cost cuts. AI can insert these lags into your formulas so forecasts behave like the real world.
• Rank levers by speed-to-cash vs size-of-cash. Fast wins first: collections calls, payment links on invoices, early-payment discounts, partial shipments with deposits.

Second prompt: collections engine (paste as-is)

“Based on this AR aging table [paste your buckets], create a weekly cash-in schedule for the next 13 weeks using these assumptions: [% collected from each bucket per week], with leakages (uncollectible %) and promised-payment dates. Output as a table I can paste into my sheet with Week and Cash In by bucket, plus a total. Include a sensitivity toggle for AR Days +/- 5 and +/- 10 that shifts receipts across weeks.”

What to expect when you run this

• A 13-week view with a clear breach week and a focus list of 3–5 levers to avoid it.
• Sensitivity to AR days and revenue shocks, so you see which move matters most.
• A light but realistic cadence: weekly updates, monthly recalibration, quarterly stress and upside tests.

Worked example (simple numbers)

• Opening Cash: 250,000; Min Threshold: 120,000.
• Weekly Cash In: 90,000 base; Weekly Out: 110,000 (two heavy payroll weeks per month).
• Base shows breach in Week 6. Applying two levers: collect 60,000 of 60–90 day AR over Weeks 2–4, and defer 40,000 CAPEX to Week 12. Result: breach moves to Week 10, giving time to add a short-term financing option or run a promo.

Mistakes to avoid (and quick fixes)

1. Mixing accrual with cash. Fix: only include actual cash dates (invoice dates are not cash).
2. Assuming instant savings. Fix: add realistic notice periods and ramp-down lags.
3. Forgetting taxes or annual renewals. Fix: add a monthly tax provision and a calendar of big renewals.
4. Using one AR days number for all customers. Fix: segment by bucket or by top-10 accounts.
5. One-way stress tests. Fix: also test an upside (e.g., +10% sales) to see if growth consumes cash (inventory, delivery, support).

1-week action plan

1. Day 1: Run the 5-minute runway check. Set your Min Cash Threshold.
2. Day 2: Generate the 13-week CSV with the prompt. Paste into your sheet. Enter known inflows/outflows.
3. Day 3: Add AR aging and the collections engine. Identify the breach week by scenario.
4. Day 4: Prioritize 3 levers by speed-to-cash and size (e.g., AR calls, early-pay discounts, vendor terms).
5. Day 5–7: Execute the first two levers. Update the sheet and re-check breach week.

Optional prompt: turn outputs into decisions

“Given this 13-week cash model summary [paste key rows], rank 6 practical actions by speed-to-cash and size-of-cash, estimate timing (including lags), list risks, and draft owner-by-week tasks. Output a simple action checklist with due dates for the next 21 days.”

Closing thought

Keep the model lightweight, refresh it weekly, and let AI do the templating and math while you make the calls. Small timing shifts, made early, beat big cuts made late.

Great call-out: your focus on thresholds, auditability, and stable pseudonyms is the difference between “we tried” and “we’re defensible.” Let’s round this out with a few insider moves that make the pipeline steadier, cheaper to run, and easier to explain to stakeholders.

Context, briefly

• Regex clears the obvious. Free text, drift, and inconsistent replacements cause most incidents.
• Auditors want evidence: metrics, logs, and repeatable rules.
• Your north star: minimize false negatives (privacy risk) while preserving analytic utility (don’t wreck the data).

What you’ll need

• A 300–1,000 row sample with free text, plus 20–30 planted canaries (benign fakes).
• Tools: regex engine, a small NER/LLM, a simple review spreadsheet, and a separate encrypted store for keys and linkage maps.
• Decision table: what gets redacted vs. generalized vs. pseudonymized (see below).

The missing piece: generalization policy (saves utility)

• Dates: keep year only or shift by a deterministic per-person offset (e.g., hash-based ±14 days). Keeps seasonality without leaking exact dates.
• Ages: convert to bands (e.g., 0–4, 5–9, …, 85+). Avoid exact over-89.
• Locations: replace full address with city or region; for rare geos, go one level broader.
• IDs/names: salted HMAC pseudonyms for longitudinal analysis, else tokenize fully.

Step-by-step (do this now)

1. Declare tiers and actions. Tier 1 (always redact or pseudonymize): names, emails, phones, national IDs, full addresses, DOB. Tier 2 (contextual): organizations, fine-grained locations, rare identifiers. Map each category to an action: REDACT, PSEUDONYMIZE, or GENERALIZE.
2. Run deterministic rules first (tested patterns). Replace with category tokens. Example patterns to copy:
   • Email: /([A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+.[A-Za-z]{2,})/g
   • Phone (broad): /(+?d{1,3}[s-]?)?(?:(?d{2,4})?[s-]?)?d{3,4}[s-]?d{3,4}/g
   • US SSN: /(?
   • Simple dates: /(?
3. Model pass on free text with conservative thresholds. Anything borderline becomes NEEDS_REVIEW. Keep span-level metadata (start, end, category, model version).
4. Apply generalization/pseudonym rules. Dates→year or shifted date; ages→bands; names/IDs→salted HMAC (keys stored separately). Ensure consistent outputs across rows.
5. Human review, risk-based. 100% review on records still containing Tier 1 after ML; 10–20% stratified sampling otherwise, plus everything marked NEEDS_REVIEW.
6. Dual-model “attacker” pass. Probe the redacted text for leaks (direct and indirect). Anything found becomes training/tuning feedback.
7. Quality gates before release. FN below threshold by category, canary detection 100%, token density reasonable (to protect utility), and a stable redaction check (identical input → identical output).

Copy-paste AI prompt (extraction + action)

“You are a compliance-grade PII redaction agent. From the input text, detect spans for: NAME, EMAIL, PHONE, ADDRESS, DATE, DATE_OF_BIRTH, NATIONAL_ID, GEO_LOCATION, ORG, ACCOUNT_ID, OTHER_PII. For each span, decide action: REDACT, PSEUDONYMIZE, or GENERALIZE. For GENERALIZE, suggest a rule (e.g., DATE→year-only, AGE→band). Return valid JSON: {spans: [{start, end, text, category, action, generalize_rule, confidence, review_flag}], redacted_text}. Rules: (1) If confidence < 0.8 set review_flag=true; (2) Preserve punctuation/spacing in redacted_text; (3) Replace spans with tokens like [NAME] or [DATE:YEAR]; (4) Do not invent information; (5) If uncertain about category or action, mark review_flag=true and choose REDACT.”

Insider trick: deterministic date shifting that preserves timelines

• Compute a per-subject offset = (HMAC_SHA256(salt, subject_id) mod 29) – 14 days.
• Shift all that subject’s dates by the same offset. Store salt separately. Result: analytic patterns survive; exact dates stay hidden.

Example (what “good” looks like)

• Original: “Spoke with John Carter on 03/14/2023 at 555-914-2231 about follow-up in 2 weeks at 14 Pine St, Boston.”
• Output: “Spoke with [NAME] on [DATE:YEAR] at [PHONE] about follow-up in 2 weeks at [ADDRESS_CITY], [ADDRESS_REGION].”
• If pseudonymizing names: “Spoke with [NAME_a1f3] …” across all rows.

Mistakes and quick fixes

• Missed PII in PDFs/images. Fix: OCR to text first, then run the same pipeline; verify with canaries embedded in images.
• Regex that’s too greedy or locale-blind. Fix: normalize text (NFKC), add locale dictionaries, and unit test patterns on a curated edge-case set.
• Over-redaction destroys joins. Fix: use category tokens and salted pseudonyms; measure token density and correlation drift vs. raw.
• Storing keys with data. Fix: separate, encrypted key store with rotation; log access.
• Confidence scores treated as truth. Fix: validate thresholds on a gold set; escalate borderline to human review.

What to expect

• Deterministic pass clears 50–70% of PII immediately.
• Model pass picks up most of the rest; plan for 5–15% manual review at the start.
• Generalization/pseudonyms preserve longitudinal and group analyses with minimal rework for analysts.

Action plan (fast track)

• Hour 1: Implement the regex pass and log the matches; plant 20 canaries and verify 100% catch.
• Hour 2: Run the extraction prompt on free text; export spans to a sheet; mark NEEDS_REVIEW.
• Hour 3: Apply generalization (year-only, age bands) and salted pseudonyms; produce redacted_text.
• Hour 4: Dual-model attacker pass; fix any misses; rerun until canaries = 100% and FN under threshold on a 300-row gold set.
• End of Day: Ship a one-page summary: precision/recall by category, FN ceiling met, token density, review rate, and sample logs.

Closing thought

Automate the heavy lift, measure what matters, and keep a human hand on the tiller until your evidence says otherwise. That’s how you get speed without surprises.

Onwards — Jeff