Jeff Bullas

Quick win (5 minutes): I love your suggestion — start with one measurable line and turn it into a benefit-first headline. Do that now: open the spec, find a measurable phrase, and rewrite it as what the customer would say out loud.

Here’s a small process to make that quick win repeatable and AI-friendly.

What you’ll need

• Product spec (one paragraph or a single feature line is enough to start).
• Buyer persona note (one sentence: main pain and decision trigger).
• Tone example (one short sample of copy you like).
• Access to an LLM (ChatGPT or equivalent) and a subject-matter expert for quick fact-checks.

Step-by-step (do this now)

1. Find one measurable spec line (e.g., “reduces sync time by 40%” or “99.95% uptime”).
2. Rewrite it as a customer sentence: benefit-first, plain language. That’s your headline candidate.
3. Use the AI prompt below to expand into three headlines, a 50-word blurb, and a 150-word feature→benefit paragraph.
4. Quick-check facts with an engineer — mark any numbers you can’t verify and remove them.
5. Put the headline live in an email or landing page and A/B test it for 3–7 days.

Copy-paste AI prompt (primary)

“You are a senior B2B product copywriter. Convert the following technical specification into marketing copy for [buyer persona: e.g., IT managers at mid-market SaaS companies]. Produce: 1) three headline options, 2) a 50-word elevator blurb, 3) a 150-word feature-benefit section that explains why it matters to the buyer, and 4) two CTAs. Use plain language, avoid technical jargon, and include one measurable benefit. Technical spec: [paste spec]. Tone: confident, helpful, concise.”

Example (from spec to copy)

Spec line: “Database sync reduced average latency by 40% under peak loads.”

• Headline candidate: “Sync your data 40% faster during peak times — no outages.”
• 50-word blurb: “Keep apps responsive when traffic spikes. Our optimized sync cuts latency by 40% under peak loads so users stay productive and support tickets drop. Easy to deploy, works with your current stack.”
• 150-word feature→benefit (trimmed): AI-generated copy that explains why lower latency matters to ops, sales, and end-users — include an example saving X minutes per task and reduced support calls.

Common mistakes & fixes

• Hallucinated numbers — fix: never publish a metric without cross-checking the spec or an engineer.
• Feature-first copy — fix: lead with the buyer’s gain, then back with the tech detail.
• One-size-fits-all tone — fix: create two versions (technical and business) and test which converts.

7-day action plan (lean)

1. Day 1: Pick 1 spec line and persona; create headline candidate.
2. Day 2: Run AI prompt, generate variants.
3. Day 3: Quick fact-check and edit.
4. Day 4: Stakeholder review and prepare two variants for A/B test.
5. Day 5–7: Run test, gather CTR and conversion data, iterate.

Small, repeatable steps beat perfect one-offs. Try that 5-minute headline now and bring me the result if you want help turning it into the 150-word version.

Make your students the historians, not the spectators. Use AI to do the heavy lifting (summaries, vocabulary, leads) — then force everything back to the document with quotes and one trusted secondary source. Fast ideation + strict verification = speed without shortcuts.

• Do keep the scan, the clean transcript, and basic metadata. Work only from your corrected text.
• Do ask the AI for short, structured outputs you can verify (claims, quotes, confidence, leads).
• Do require an anchor quote for every AI claim. No quote, no claim.
• Don’t accept AI attributions, dates, or provenance without archival or scholarly confirmation.
• Don’t put student-identifiable information into public tools.

What you’ll need

• Document scan (image/PDF) and a corrected transcript.
• Metadata: author (if known), date, place, and archive path.
• School‑approved AI tool (or offline) and a folder/LMS to save outputs.
• A simple “evidence ledger” template students can fill in.

The 3‑pass routine (30–45 minutes total)

1. Pass 1 — Triage (10–15 min): Ask AI for a neutral 1‑paragraph summary, named entities, three unfamiliar terms, and the top five claims it believes the author is making — each with a one‑sentence justification and a confidence tag (H/M/L).
2. Pass 2 — Bias & context (10–15 min): Ask AI for three likely perspectives or biases in the text, what might be missing (voices or data), and three practical leads to corroborate or refute the claims (newspaper, census, price records, legislative acts).
3. Pass 3 — Verification (10–15 min): Students choose one AI claim each. They must attach an exact anchor quote from the transcript, then check a secondary source or another primary item. Record verdict: supported, contradicted, or unresolved — with a citation/quote.

Copy‑paste prompts (robust and classroom‑ready)

Prompt 1 — Triage with anchors You are a cautious historical research assistant. Analyze the transcription and metadata below. Produce sections: (1) Neutral summary (≤120 words). (2) Named people, places, dates. (3) Three unfamiliar or era‑specific terms with plain definitions and why they matter. (4) Top five claims the author appears to make — for each: the claim in one sentence; the exact anchor quote from the transcript with line number or surrounding words; confidence (high/medium/low) with one‑sentence reasoning; one verification step a student could do. If you cannot find a direct quote for a claim, label it “NO ANCHOR — DROP”. Do not invent facts or dates.

Prompt 2 — Bias, omissions, leads From the same text, list: (1) Three likely perspectives or biases visible in the writing and the lines that signal them. (2) What’s missing — voices, regions, data — that could skew interpretation. (3) Three concrete leads to corroborate or challenge the claims (name the type of record and the exact detail to look for). Flag any uncertainties.

Prompt 3 — Cross‑source check Compare Primary A (the transcript) with Secondary B (a vetted article or textbook excerpt). Output: agreements with anchor quotes from A and page/section from B; contradictions with both sources quoted; silent spots where B does not discuss A’s key points; a short list of next sources to consult. Keep judgments cautious and tied to quotes.

Evidence ledger (student template)

• Claim ID (from AI):
• Anchor quote (exact words + line number):
• AI confidence (H/M/L) + reason:
• Student verdict (supported/contradicted/unresolved):
• Verification source + exact citation/quote:
• Notes (bias, missing voices, OCR concerns):

Worked example (suffrage pamphlet, c. 1913)

• Pass 1: AI summary notes a call for “lawful reform,” names a city, cites “property‑owning women,” flags unfamiliar terms like “cat and mouse” (era‑specific policy). Claims include: taxation without representation harms women; moral authority improves governance. Each claim comes with an anchor quote and confidence tag.
• Pass 2: Biases flagged: middle‑class perspective; legalist strategy; omission of working‑class conditions. Leads: local newspaper coverage of a march (date given in text), tax roll records for women property holders, parliamentary debates from that session.
• Pass 3: Students verify: one checks a newspaper archive for the march; another checks a legislative record for the debated bill. Each logs a verdict with citations.

Insider tricks that lift quality

• Anchor‑or‑drop rule: Any AI claim without a verbatim quote is removed from discussion.
• Opposition reading: Ask the AI to draft a 3‑point counter‑argument a contemporary critic might make, each tied to a line in the text. Students test which counters the source anticipates or ignores.
• Time‑shift vocabulary: Have AI list words likely to have shifted meaning (“liberty,” “rate,” “corn”) with period‑specific definitions and the lines where they appear.
• Claim budget: Cap the AI at five claims. Scarcity forces clarity and makes verification manageable in one class.

Common mistakes and quick fixes

• Mistake: OCR errors skew quotes. Fix: proof the transcript first; if in doubt, paste a short image snippet and ask AI to list uncertain words.
• Mistake: Students copy AI phrasing. Fix: grade the ledger (quotes + citations), not the AI text.
• Mistake: AI invents provenance. Fix: require a source type and an anchor quote for any claim about origin or date; otherwise mark “unverified.”
• Mistake: Too many leads. Fix: one lead per student, five total per document.

1‑week rollout (measurable)

1. Pick one document. Save scan + corrected transcript + metadata.
2. Run Prompt 1 and Prompt 2. Print or post outputs.
3. In class: students annotate and fill the evidence ledger for one claim each.
4. Homework: verify with one secondary or corroborating primary source; capture exact citation/quote.
5. Next class: share verdicts; compute accuracy rate (supported claims ÷ total claims) and time saved vs. a prior, non‑AI lesson.

What to expect: quicker entry to the text (often 30–50% faster triage), richer discussion of bias, and a visible trail from claim → quote → citation. You’ll still see confident AI mistakes — the anchor‑or‑drop rule keeps them contained.

Bottom line: let AI accelerate discovery, but you and your students own proof. If it isn’t anchored to the document and checked once, it doesn’t fly.

Agreed: your tokenized skeleton + reconciliation + change‑order flow is the sweet spot. Let’s add three upgrades that reduce rework and win deals faster: a single “parameter block” you paste once, Good–Better–Best options from the same pricing sheet, and a quick red‑team pass that hunts risk before a client does.

Why this helps: One source of truth is great. A parameter block makes it portable across tools. Options increase acceptance without new typing. And a red‑team pass catches ambiguity, unbounded scope, and payment risks in minutes.

What you’ll need

• Your pricing sheet with Row IDs (authority).
• Tokenized SOW skeleton (you have it).
• The parameter block template (below).
• Optional: a simple holiday list or “business days only” note.
• Access to an AI chat writer.

Step‑by‑step (fast and safe)

1. Create a portable parameter block you can paste into any AI chat. It mirrors your sheet. Keep all numbers and dates here. If something’s missing, the AI must stop and ask.
2. Draft pass: feed the brief, Row ID, and the parameter block. Tell the AI to fill your tokenized skeleton and to ignore any number not present in the block.
3. Options pass (Good–Better–Best): from the same Row ID(s), ask the AI to produce three clearly labeled packages: Essential, Standard, Premium. Each with scoped differences, revision caps, and price. Attach the chosen option as Appendix A.
4. Reconciliation pass: list every number and date in the SOW and match it to the parameter block/Row ID. Flag mismatches with a one‑line fix.
5. Red‑team pass: have the AI act as client CFO + project manager. It should hunt for open‑ended scope, vague acceptance, weak payment terms, and date risks (weekends, missing client lead times).
6. Stamp and send: insert a version stamp (Doc ID + Row ID + date/time). Send as an editable PDF with Appendix A pricing snapshot.

Parameter block template (copy and fill)

PROJECT: Website redesign for [[CLIENT]] to deliver [[OUTCOME]] by [[TARGET_DATE]]
ROW_ID: [[ROW_ID]]
MODEL: Fixed‑Fee | Time & Materials | Retainer
CURRENCY: [[CURRENCY]]
TAX_RATE_OR_VALUE: [[TAX]]
RATES: [[ROLE_1]]=[[RATE_1]]; [[ROLE_2]]=[[RATE_2]]
EST_HOURS: [[HOURS_TOTAL]] (or leave blank for fixed fee)
PRICING: Subtotal=[[SUBTOTAL]]; Tax=[[TAX_VALUE]]; Total=[[TOTAL]]; Terms=[[TERMS]]
MILESTONES: KO=[[DATE_KO]]; M1=[[DATE_A]]; M2=[[DATE_B]]; Final=[[DATE_FINAL]]
DELIVERABLES: [[DELIVERABLE_1]] (accept=[[ACCEPT_1]]); [[DELIVERABLE_2]] (accept=[[ACCEPT_2]])
IN_SCOPE: [[IN_SCOPE_BOUNDARIES]]
OUT_OF_SCOPE: [[OUT_OF_SCOPE]]
CLIENT_RESPONSIBILITIES: Feedback within [[DAYS]] business days; provide assets by [[DATE_ASSETS]]
REVISION_CAP: [[REVISION_ROUNDS]] rounds per deliverable; extra at [[REV_RATE]]/hr
BUSINESS_DAYS_ONLY: Yes

Copy‑paste AI prompt: Master SOW package (draft + options + checks)

You are my SOW compiler. Use ONLY the values in the Parameter Block and the Pricing Row ID. If any value is missing or inconsistent, stop and ask for it before drafting. Tasks: 1) Fill the tokenized 1‑page SOW skeleton; 2) Generate Good–Better–Best options from the same source (Essential, Standard, Premium) with clear scope differences, revision caps, and exact pricing tied to Row ID values; 3) Produce a reconciliation list of every number and date in the SOW vs. the Parameter Block with pass/fail and a one‑line fix; 4) Run a red‑team critique as a client CFO and project manager, listing risks (open‑ended scope, vague acceptance, weak payment terms, weekend/holiday dates, missing client lead times); 5) Output a Version Stamp with Doc_ID=[[CLIENT]]‑[[ROW_ID]]‑v1 and timestamp. Use plain English. Do not invent numbers. If MODEL is T&M or Retainer, format Costs & Terms accordingly.

How the options should look

• Essential: core deliverables only, conservative hours, tight acceptance, lowest price.
• Standard: Essential + 1–2 extras (e.g., training, extra templates), moderate revision cap.
• Premium: Standard + strategy workshop or support period, highest revision cap, priority response times.

Mini example (inputs you paste)

• Brief: Redesign a 10‑page website; launch by April 30.
• Parameter Block: use the template above with Row ID 104 and the dates/rates you listed earlier.
• Ask for: SOW + Options + Reconciliation + Red‑team + Version Stamp.

Mistakes & quick fixes

• AI “helpfully” adjusts totals — Fix: instruct “never calculate totals; use PRICING from the Parameter Block exactly. If missing, ask.”
• Weekend or holiday deadlines — Fix: include BUSINESS_DAYS_ONLY=Yes; ask the AI to bump to next business day and note the change.
• Option bloat — Fix: cap each option to a named set of deliverables and a revision cap. Everything else is a change order.
• No audit trail — Fix: embed Version Stamp and attach Appendix A (pricing snapshot of Row ID).

What to expect

• Draft + Options in 3–6 minutes once your Parameter Block is ready.
• Reconciliation and red‑team output in 2–5 minutes.
• A single human pass focused on numbers, dates, and client responsibilities (5–10 minutes).

10‑minute action plan

1. Copy the Parameter Block template and fill it for one live deal.
2. Paste your tokenized skeleton + filled Parameter Block + Master SOW prompt into your AI.
3. Pick an option (Essential/Standard/Premium), attach the pricing snapshot, and insert the Version Stamp.
4. Run the red‑team list and fix any flagged items before sending.

Insider tip: Most delays aren’t writing—they’re decisions. The options pattern helps clients choose fast without another meeting. The parameter block keeps your math honest. Together, you get fast drafts, clean numbers, and fewer surprises.

Spot on: making the pricing sheet the authority is the move. Now let’s bolt on three add‑ons that cut errors further and shave more minutes: a tokenized SOW skeleton, an AI reconciliation check, and a one‑click change‑order flow.

The idea: keep numbers and dates outside the draft (single source of truth), then force the AI to 1) fill a standardized SOW, 2) produce a cross‑check list, and 3) generate a clean change order when anything shifts.

What you’ll need

• Pricing sheet with Row IDs (your authority).
• One‑paragraph brief and deliverables with acceptance criteria.
• A tokenized SOW skeleton (below) in your template tool.
• Access to an AI chat writer.

Step‑by‑step

1. Create a tokenized SOW skeleton once. Use simple tokens the AI can fill from your pricing row and brief.
2. Draft pass: feed the brief, Row ID, and deliverables to the AI with the “SOW Draft” prompt (yours works great).
3. Reconciliation pass: ask the AI to list every number/date in the draft and confirm it matches Row ID X exactly. Fix any mismatches.
4. Change‑order ready: when scope or dates change, run the “Delta” prompt to generate a clean change order without rewriting the SOW.
5. Final polish: run a readability pass (grade 7–9, plain English) and a quick manual checklist on rates, dates, and responsibilities.

Premium template: tokenized 1‑page SOW skeleton

• Overview: Project for [[CLIENT]] to deliver [[OUTCOME]] by [[TARGET_DATE]].
• Scope: We will deliver [[SCOPE_ITEMS]] within [[IN_SCOPE_BOUNDARIES]]. Out of scope: [[OUT_OF_SCOPE]].
• Deliverables: [[DELIVERABLE_1]] — acceptance: [[ACCEPT_1]]. [[DELIVERABLE_2]] — acceptance: [[ACCEPT_2]].
• Timeline & Milestones: Kickoff [[DATE_KO]], Milestone A [[DATE_A]], Milestone B [[DATE_B]], Final [[DATE_FINAL]].
• Costs & Payment: Pricing from Row ID [[ROW_ID]]: Rate [[RATE]] x Qty [[QTY]] = Subtotal [[SUBTOTAL]]; Tax [[TAX]]; Total [[TOTAL]]. Terms: [[TERMS]]. Currency: [[CURRENCY]].
• Assumptions: [[ASSUMPTIONS]].
• Client Responsibilities: [[CLIENT_RESPONSIBILITIES]].
• Change Control: Requests outside scope trigger a written change order with updated price and dates before work proceeds.
• Acceptance: Work is accepted when deliverables meet the criteria listed above and are approved in writing within [[ACCEPT_DAYS]] days.
• Appendix A: Pricing snapshot from Row ID [[ROW_ID]].

Copy‑paste AI prompt: SOW reconciliation checklist

Compare the SOW I will paste with the single source of truth: Pricing Row ID [[ROW_ID]] and milestone dates. Create a checklist with: 1) every number and date in the SOW, 2) the corresponding value from Row ID [[ROW_ID]], 3) a pass/fail flag, and 4) a one‑line fix when failed. Do not rewrite the SOW. Output only the checklist.

Copy‑paste AI prompt: change‑order (delta‑only) generator

You are a change‑order writer. I will provide: the signed SOW summary and a list of changes. Produce a one‑page Change Order with sections: Summary of Change, Impact on Scope, Revised Milestones (dates only), Revised Costs (use pricing from Row ID [[ROW_ID]] exactly), Assumptions, and Signature. Keep plain English. List ONLY what changed. If any input is missing (rates, dates, acceptance impact), ask precise questions before drafting.

High‑value extras

• Pricing model toggles: In your prompt, declare one of these before drafting: “Model: Fixed‑Fee,” “Model: Time & Materials,” or “Model: Retainer.” The AI will format Costs & Terms correctly.
• Revision caps: Add a standard line to avoid scope creep: “Includes up to [[REVISION_ROUNDS]] revision rounds per deliverable; extra rounds billed at [[REV_RATE]].”
• Acceptance patterns (copy into intake): “Homepage accepted when: a) loads under 3s on desktop, b) passes QA checklist, c) approved by email.” Make acceptance measurable, not descriptive.
• Currency and tax guardrail: Include explicit tokens for currency and tax source. If missing, the AI must flag it as a blocker before drafting totals.

Mini example (what goes in)

• Brief: Redesign company website with 10 pages; launch by April 30.
• Row ID 104: Rate 120/hour; Est. 120 hours; Subtotal 14,400; Tax 1,440; Total 15,840; Terms 50/50; Currency USD; Milestones: KO Mar 5, Design Mar 20, Build Apr 10, Launch Apr 30.
• Deliverables: IA + wireframes; 10 page templates; CMS setup; QA + launch. Each with acceptance tests.

Common mistakes & fast fixes

• Numbers in two places — Fix: Never type totals in the draft manually; always pull from Row ID and attach Appendix A.
• Vague acceptance — Fix: Require a test or evidence (file, URL, metric, or date) for each deliverable.
• Uncapped revisions — Fix: Add revision caps and an hourly uplift for extras.
• Slippery dates — Fix: Put client response times in responsibilities (e.g., “client feedback within 3 business days”); late feedback shifts dates.

What to expect

• Draft in 2–5 minutes using your existing prompt.
• Reconciliation checklist in 1–3 minutes; human verify in 2–5 minutes.
• Change order generated in 3–6 minutes when scope or dates move.

90‑minute rollout

1. Paste the tokenized skeleton into your SOW template.
2. Pick 1 live project; assign a Row ID; fill missing tokens (currency, tax, revision cap).
3. Run Draft prompt, then Reconciliation prompt; fix mismatches.
4. Attach Appendix A pricing snapshot and send for one‑round sign‑off.
5. When a change arrives, test the Delta prompt to issue a clean change order.

Bottom line: Let the AI write the words; let your sheet own the numbers. Add reconciliation and delta prompts, and you’ll move from “fast drafts” to “fast, correct, and defensible agreements.”

Smart take — your two-pass method and four levers are the 80/20. Let’s add one more piece that turns this into a predictable, teacher-friendly workflow: a simple delta-tuner that converts “how far off” you are from the target into exact edits you can make in minutes.

Why this helps: AI can guess readability, but only your checker or service confirms it. So we draft fast, measure once, then make precise edits based on the gap. That’s how you hit a tight band without endless rewriting.

What you’ll need

• Target Lexile band (e.g., 420–480L) and word count (100–250 words).
• Topic and 4–6 target vocabulary words with parenthetical definitions.
• A readability/Lexile checker (use your preferred tool) and a 5-minute pilot with 3–5 students.
• A simple log to track: estimated Lexile, avg sentence length, longest sentence, total words, and common stumble words.

Step-by-step (adds a delta-tuner loop)

1. Plan (your Pass 1): Ask the AI to produce a sentence plan (count, length per sentence, where vocab appears, which two can be compound). Keep average 9–11 words; max 16.
2. Draft (your Pass 2): Write to the plan. Require first-use definitions and a neutral, age-respectful tone. Ask the AI to list its own stats (words, avg/longest sentence) — then you verify with your checker.
3. Measure: Run the draft through your checker. Record the score and stats in your log. Note any long sentences and rare words.
4. Delta-tune: Use the measured gap to pick precise edits:
   • If you’re high by 60–120L: split the longest sentence; remove one dependent clause; replace two rare words with target words (keep glosses); trim adverbs or stacked prepositional phrases.
   • If you’re low by 60–120L: add one compound sentence; introduce one grade-appropriate academic word with a parenthetical definition; add a comparative or causal phrase once.
   • Re-check after no more than three edits. Aim to move 60–100L per pass.
5. Scaffold: Add 2 questions (one literal, one inferential), 1 writing prompt, 3–5 term glossary. Make the questions echo the passage’s verbs and nouns to reduce construct-irrelevant difficulty.
6. Pilot: 3–5 students; capture minutes to read, words-per-minute, % correct, engagement 1–5, and stumble words. If time exceeds your target by 25% or comprehension dips under 70%, step down one band for instructional use.
7. Finalize: Make one tone edit and one vocabulary swap based on pilot notes. Re-check readability, archive the final along with its stats.

Copy-paste AI prompt: Delta‑Tuner revision

“You are revising a leveled reader passage to hit a target Lexile band. Here are the inputs:
Target band: [e.g., 420–480L]
Measured score from checker: [e.g., 560L]
Word count target: [e.g., 160–190]
Constraints: average sentence length 9–11 words; max 16 words; neutral, age-respectful tone; keep these vocabulary words with first-use definitions: [list].
Text to revise:
[PASTE PASSAGE]
Tasks: (1) Diagnose why the score is off (sentence length, rare-word density, clauses); (2) Apply up to three micro-edits using this priority: split the longest sentence; replace one rare word with a target word; remove or add one clause as needed; (3) Output the revised passage only; (4) Then output a short change log and your readability stats: total words, average sentence length, longest sentence. Do not invent a ‘true Lexile’; the checker will verify.”

Copy-paste AI prompt: One-topic, three-band mini set

“Create three versions of the same informational passage on [topic]. Bands: Low [Target−100L], Target [Target], High [Target+100L]. Constraints for all: neutral, age-respectful tone; accurate, concrete facts; 3–5 term glossary; 2 questions (literal + inferential); 1 short writing prompt. Additional constraints: Low = 110–140 words, avg sentence 8–10 words; Target = 160–190 words, avg 9–11; High = 210–250 words, avg 11–13 with 2 compound sentences. Include these vocabulary words at the Target level with first-use definitions: [list]. Repeat each target word 2–3 times across levels. End each version with estimated readability stats (not a verified Lexile).”

Mini example of delta‑tuning (how it feels in practice)

• Before (too hard): “Bats navigate by emitting ultrasonic pulses, which bounce off objects and return detailed echoes.”
• Edits: Split the sentence; swap “ultrasonic” for a target word with a gloss; remove “detailed.”
• After: “Bats send out quick sounds. The sounds bounce back as echoes. This helps them find insects.”

Common mistakes and quick fixes

• AI ‘Lexile’ is treated as official → Always verify with your checker; AI’s number is only a guide.
• Older striving readers get baby talk → Specify “mature topics, simple language, no sing-song phrasing.” Keep dignity high.
• Vocab defined once, then forgotten → Repeat target words 2–3 times. Keep definitions short and consistent.
• Questions don’t match the text → Mirror the passage’s key verbs and nouns in stems and choices.
• Early grades need decodability → If phonics-first is the goal, control sound-spelling patterns; Lexile alone won’t do it.

Fast rollout plan (90 minutes)

1. Minutes 0–15: Set your band and guardrails; pick topic and 4–6 vocab words.
2. Minutes 15–35: Run the two-pass prompts. Get the first draft with self-reported stats.
3. Minutes 35–50: Verify with your checker. Apply one delta‑tune pass (max three edits). Re-check.
4. Minutes 50–70: Add two questions, one writing prompt, and a 3–5 term glossary.
5. Minutes 70–90: Pilot with 3–5 students; log WPM, % correct, engagement, stumble words. Make two final tweaks.

Set expectations: You’ll usually land within ±75L after one delta‑tune pass. Plan on 2–3 total passes for tricky topics or when tone matters a lot. Human review catches the small factual or cultural slips AI can miss.

Pick one topic and one target band today. Run the two-pass + delta‑tuner loop once. By tomorrow, you’ll have a clean, classroom‑ready leveled reader and a repeatable system you can trust.

Nice follow-up — your playbook is tight and the ROI math is exactly what small budgets need. I’ll add a few practical shortcuts and a ready-to-use AI prompt so you can find winning keywords faster and avoid common traps.

What you’ll need

• 3–5 seed phrases for your main product or service
• Target region (city/zip), monthly ad budget, and a guessed conversion rate (2–5% is fine)
• Access to Google Keyword Planner (or a basic keyword tool) and an AI chat assistant
• A simple landing page aligned to your ad message

Step-by-step (quick, do-first play)

1. Seed + Planner: Put seeds into Keyword Planner. Note CPC range and volume for each seed.
2. Ask AI to expand: Use the prompt below to get 20–30 long-tail, buying-intent variants plus intent/CPC estimates.
3. Filter fast: Keep keywords with estimated CPC ≤ your target CPC and intent = high or medium.
4. ROI check: clicks = budget ÷ CPC; conversions = clicks × conv. rate; CPA = budget ÷ conversions. If CPA fits your margin, test it.
5. Structure: 3 tight ad groups, phrase + exact match only, add 20 obvious negatives (free, jobs, DIY, etc.).
6. Test: run with low daily caps for 7–14 days, review search terms, pause losers, double-down on winners.

Worked example (fast)

Budget $300/mo. Target CPC $1.20. Conv. rate 3% → clicks = 250, conversions ≈ 7.5 → CPA ≈ $40. If $40 fits margin, proceed; if not, improve landing page or tighten keywords (more specific/location-based).

Common mistakes & fixes

• Too many keywords: cut to top 20% performers after 10–14 days.
• No negatives: check Search Terms daily and add negatives immediately.
• Irrelevant landing page: match headline to keyword and offer before increasing bids.

Copy-paste AI prompt (use this exactly)

Prompt: “I sell [product/service] in [city/region]. Generate 25 long-tail Google Ads keywords with strong buying intent. For each keyword, label intent as High/Medium/Low, estimate a likely CPC range for a small local campaign, and score each keyword 1–10 for expected ROI assuming a 3% conversion rate. Also suggest 5 negative keywords to block irrelevant traffic.”

7-day action plan

1. Day 1: Run Keyword Planner on 3–5 seeds and paste into AI prompt.
2. Day 2: Filter to 15–25 keywords, create 3 ad groups, add negatives, write 3 ads per group.
3. Days 3–7: Run with daily caps, review search terms, pause irrelevant keywords, boost bids on early winners.

Reminder: With a small budget, speed beats perfection — find one winner, then scale. Keep testing headlines and landing page tweaks while the ads run.

Quick win: In under 5 minutes, run an auto-transcription of a short audio clip (Whisper-like tool) and ask an AI to give you three action-first bullets from that transcript.

Small correction before we start: AI won’t magically understand mixed inputs without a bit of prep. Audio needs transcription, images often need OCR or scene descriptions, and timestamps or metadata help tie everything together. With that in mind, here’s a practical approach you can try.

What you’ll need

• An audio transcription tool (Whisper, Otter, or built-in service).
• An OCR/image description tool (Tesseract, or models that accept images).
• A multimodal summarizer or a text-based LLM (GPT-like) to combine the extracted text.
• A simple workflow tool or folder to collect files and timestamps.

Step-by-step

1. Collect inputs: gather your text files, audio recordings, and images in one place.
2. Transcribe audio: convert audio to text and keep timestamps for important parts.
3. Extract from images: run OCR for text in images and/or request a short scene description for photos or slides.
4. Normalize and tag: add simple tags (topic, speaker, time) so pieces can be aligned.
5. Merge into a single document: combine transcripts, image text, and notes in chronological or thematic order.
6. Ask the AI to summarize: request a concise summary, key insights, and action items.
7. Review and refine: check for errors, add context, and prioritize the actions.

Copy-paste AI prompt (use after you’ve combined the text):

Take the following combined material (transcript snippets with timestamps, image text, and notes). Produce a short clear summary: 3 key insights, 4 recommended actions ranked by priority, and any items that need clarification. Keep each insight to one sentence and actions to one line each. Include references to timestamps or image captions when relevant.

Example

Inputs: meeting audio (0:02:15 – vendor concern), slide image with sales chart (OCR text: Q3 up 12%), and chat notes. Result: 1) Sales rising Q3 +12% (slide); 2) Vendor delay risk at 0:02:15 — consider alternative supplier; 3) Need clearer KPI dashboard. Actions: 1) Contact backup supplier (high), 2) Update dashboard spec (medium), 3) Share summary with team (low).

Common mistakes & fixes

• Noisy audio → use noise-reduction or ask for a short re-recording.
• Unreadable images → get higher resolution or manually transcribe key text.
• Too much duplicated content → dedupe before summarizing.
• Blind trust in AI → always spot-check facts and timestamps.

Simple action plan (next 30 minutes)

1. Pick one 3–5 minute audio clip and one image.
2. Transcribe the audio and run OCR on the image.
3. Paste the combined text into the AI using the prompt above.
4. Review the output, pick one action and do it.

Start small, repeat, and you’ll get faster at turning mixed inputs into clear, useful insights.

Smart additions: your metrics and the 3-variant discipline are spot on. Let’s bolt on the “last-mile” production pass and a reusable template so your mockups scale across sizes and channels without rework.

Do / Do not (production pass)

• Do: Keep one master layout, then export derivatives (desktop, mobile, ad sizes) from that master.
• Do: Use vector for logos/icons (SVG), photos/illustrations as WebP (PNG only if you need transparency).
• Do: Name files like 2025-01_brand_hero-desktop_v1.webp; bump versions only when specs change.
• Do: Budget file size (goal: hero ≤180KB WebP, thumbnails ≤60KB).
• Do not: Bake text into images unless it’s a logo—keep copy editable for A/B tests and accessibility.
• Do not: Export one scale; always 1x and 2x (e.g., 1600×600 and 3200×1200).

What you’ll need

• AI image tool for mockups/variants
• Figma or Canva (vector-friendly)
• Simple contrast checker (target 4.5:1 body, 3:1 large text)
• Staging preview (no-code page or simple HTML)
• Export checklist and folder template

Insider trick (saves hours): Build a single Master Frame with styles (colors, text, spacing). Duplicate that frame to auto-generate all sizes. When you tweak the master’s color or font, all derivatives update. Then export everything in one click. Expect 20–30 minutes to set up; it repays every project.

Step-by-step: the 45-minute last-mile production pass

1. Set up the master (10 min): In Figma/Canva, recreate the chosen design with Auto Layout or grouped layers. Define color styles (HEXs), text styles (H1/H2/body), and an 8px spacing scale.
2. Create derivatives (10 min): Duplicate the master for key sizes: Desktop hero (1600×600), Mobile hero (390×480), Social square (1080×1080), Portrait (1080×1350), Landscape ad (1200×628). Keep gutters and safe areas consistent.
3. Accessibility pass (5 min): Check contrast ratios. Add alt text notes using this formula: role + what + context + brand tone (e.g., “Illustration of confident professional reviewing finances; calm, trustworthy tone”).
4. Export (10 min): Logos/icons as SVG. Photos/illustrations as WebP (1x and 2x). Keep PNG only where you need transparency. Record final sizes in your README.
5. Preview + performance (5–10 min): Drop assets into staging. Confirm rendering, crispness at 100% zoom, and total KB added. Aim for ≤250KB added above the fold.

Folder template (copy this)

• /project_name/
  • /masters (Figma/Canva + SVGs)
  • /exports/
    • /webp (1x, 2x)
    • /svg
  • /spec (PDF style sheet, tokens, README)
  • /previews (staging screenshots)

Worked example: extend your hero into a full, shippable set

1. Master hero (desktop): 1600×600. H1 40px/120% line-height, subhead 20px, button 16px. Colors: #0B61A4 (primary), #F5EFE6 (background), #0E2233 (text). Spacing scale: 8/16/24/32.
2. Derivatives:
   • Mobile hero: 390×480. Stack elements vertically; increase H1 to 44px for legibility at small widths.
   • Social square: 1080×1080. Keep text safe area central 70%.
   • Portrait: 1080×1350. Shift CTA above the fold line (upper two-thirds).
   • Landscape ad: 1200×628. Shorten headline to 6–8 words.
3. Exports: SVG logo; hero-photo.webp (1600×600, 3200×1200), hero-illustration.webp (same sizes); mobile-hero.webp (390×480, 780×960); social.webp (1080, 2160); portrait.webp (1080×1350, 2160×2700); landscape.webp (1200×628, 2400×1256).
4. Alt text samples: “Photo of professional reviewing a budget on a laptop; clear call-to-action to book a coaching call.” “Flat illustration of secure savings growth; blue and warm neutral palette, welcoming tone.”
5. QA: Open on your phone at 100% zoom. Print the hero at A4 to catch spacing/contrast issues fast. Fix, re-export, done.

Robust, copy-paste AI prompts

• Variant generation: “Create 10 layout variations for a website hero promoting financial coaching for professionals 40+. Provide both photo-led and illustration-led options. Reserve space for a 6–8 word headline, 1-line subhead, and a primary button. Use trust blues and warm neutrals. Output clean compositions that can be rebuilt as vectors. Show desktop (1600×600) and mobile (390×480) crops.”
• Token extractor: “From this image, propose a production style set: 3 brand colors with HEX, text styles (H1/H2/body with pixel sizes and line-height), and spacing tokens on an 8px scale. Return as a concise list I can paste into Figma styles.”
• Alt text maker: “Write 1 sentence of descriptive, non-marketing alt text for each asset. Use this structure: role + what + context + tone. Keep under 120 characters.”

Common mistakes & quick fixes

• Crowded edges: Text touches borders in smaller crops — add a 24px safe margin (mobile), 32px (desktop).
• Soft logos: Raster logos look fuzzy — rebuild as SVG once, then never touch again.
• Oversized files: Hero >200KB — switch to WebP, reduce background texture, and limit color noise in photos.
• Inconsistent CTAs: Button sizes shift across exports — lock a button component tied to your text styles.

Action plan (3 focused sessions)

1. Session 1 (60–90 min): Generate 10–12 variants with the prompt. Triage 3 winners. Rebuild the master frame with styles.
2. Session 2 (60 min): Duplicate derivatives (desktop, mobile, 3 ad sizes). Run contrast and alt text. Export SVG/WebP at 1x and 2x.
3. Session 3 (30–45 min): Stage, measure KB and LCP impact, tweak, and zip the bundle with README and spec PDF.

What to expect: A clean, versioned bundle that ships in under 48 hours: 1 master, 5 derivatives, vector logos, alt text library, and a one-page style sheet. Two quick revision cycles, then launch your A/B with confidence.

Closing nudge: Don’t aim for perfect art — aim for consistent specs. Ship the first bundle, test one variable, and let the data buy the next iteration.