AI Hands-on Training · From the numbers to a 30-day plan
The full six-hour hands-on training: first get the numbers right (tokens vs. labor), then learn to state requirements (the four-element method), then put tools to work. After that, walk through 19 real workplace needs collected from a company learning group — sales, R&D, HR and QA, common across pharma trading peers. The focus is solution thinking: methods, paths and pitfalls. The final unit sends you away with a 30-day plan.
Agenda (10:30–17:30)
Net learning time is about six hours after lunch and tea breaks. Click a unit to jump to it.
| Time | Length | Topic |
|---|---|---|
| 10:30–11:00 | 30 min | Opening + Lesson 1: the AI bill — tokens vs. labor → |
| 11:00–12:00 | 60 min | Lesson 2: say it right — prompts and requirements → |
| 12:00–13:00 | — | Lunch break |
| 13:00–14:00 | 60 min | Lesson 3: tools you can try today → |
| 14:00–15:15 | 75 min | Lesson 4: 19 questions, part I — sales & business (10) → |
| 15:15–15:30 | 15 min | Tea break |
| 15:30–16:30 | 60 min | Lesson 5: 19 questions, part II — R&D (6), HR (2), QA (1) → |
| 16:30–17:30 | 60 min | Lesson 6: workshop, the 30-day plan and live Q&A → |
Three learning goals
Explain how token cost compares with labor cost, and when AI is — and is not — worth using.
Master the four elements — context, constraints, format, audience — to get usable output on the first try.
Walk out with a three-step path and a 30-day plan for one real scenario in your own team.
The AI bill: tokens vs. labor
Tokens are petty cash; people's hours are the real money- · Explain what a token is and how billing works
- · Judge AI economics by cost per unit of output
- · Spot four situations where AI must not deliver the conclusion
1. Tokens in plain language
The character-ish unit AI services bill by — like mobile data billed in MB. One Chinese character is roughly 1–2 tokens; a multi-thousand-character report costs tens of thousands to hundreds of thousands of tokens. Pay as you go.
Rule of thumb: a typical office task costs pennies to a few yuan in tokens, while the human time it replaces is measured in hours. Tokens are the water bill; people's time is the payroll. Denying people water to save the water bill is bad bookkeeping.
2. Five takeaways
- 1.Token cost is usually a fraction of labor cost: a task costs pennies to a few yuan, while the same output by hand takes hours. What you save is hours, not tokens.
- 2.Unclear requirements are the biggest waste: rework burns expensive human hours, while tokens stay cheap. Learning to ask beats buying more quota.
- 3.Manage tokens, don't hoard them: set a budget and review where it goes — but never trade hours of human work to save a few yuan.
- 4.Judge cost per unit of output, not per call: one good prompt template serves today, tomorrow and the whole team — the marginal cost of repeated work approaches zero.
- 5.Human value moves up the stack: AI drafts, humans decide. Judgment, sign-off and accountability stay with people — the stronger AI gets, the more people matter.
3. A one-page bill
| Task | Manual | With AI | What you save |
|---|---|---|---|
| Market research draft | 4 h × hourly rate | 1 h × rate + a few yuan of tokens | ≈ 3 person-hours |
| Weekly email / report digest | 1 h every week | 10 min + pennies of tokens | ≈ 50 min every week |
| Training deck draft | 3 h | 1 h + a few yuan of tokens | ≈ 2 person-hours |
| Repeated Q&A (policy reading etc.) | Constant interruptions, adding up | Knowledge-base assistant answers; humans cover edge cases | Fragmented time reclaimed |
How to read it: the token cost of all four rows is less than a coffee, yet they save a dozen person-hours. Conversely, unclear requirements that force three rounds of rework waste human supervision time — which is why Lesson 2 follows immediately.
4. When not to use AI
- ·Unmasked confidential data: customer lists, pricing, formulas and unpublished processes never enter public models (see Lesson 3, data red lines).
- ·Numbers that must be exactly right: HS codes, quotations, regulatory conclusions — AI errs, and a wrong answer is worse than a missing one. Verify before use.
- ·Judgments with sign-off responsibility: picking suppliers, setting prices, issuing compliance conclusions — AI analyzes, humans sign.
- ·Trivial one-off tasks: if a single sentence describes it, just do it — faster than prompting.
5. Where the bill shows up in the 19 questions
| Question | Cost pain | How the math works |
|---|---|---|
| #6 project info gaps | Hours spent reporting progress instead of making it | AI gathering changes buys fragmented time back in blocks |
| #12 unclear requirements | Output nothing like what was wanted | Rework costs twice: human waiting plus duplicate tokens. The four elements pay for themselves |
| #14 impurity study grading | Over-study wastes materials, people and time | AI-assisted grading trades small token spend for big lab savings |
| #15 method development | Aimless trial-and-error runs | Predict first, test second — fewer failed runs |
Unit quiz (instant score + explanations)
Lesson 1 quiz: five questions, instant score with per-question explanations. Nothing is uploaded — results stay in your browser.
Your score stays in this browser only — nothing is uploaded or recorded.
Say it right: prompts and requirements
Treat AI as a smart, eager colleague on their first day- · State requirements once, using the four elements
- · Know what to give up front and what to iterate
- · Control output with hallucination drills and the three credibility checks
1. The four elements
A workable prompt = context + constraints + format + audience. Whatever you leave out, AI fills in on its own.
| Element | Question to answer | Weak (AI must guess) | Strong (AI has ground to stand on) |
|---|---|---|---|
| Context | Who are you? What is the backstory? | 'Write a market analysis' | 'I am a sales manager at an API exporter, convincing an Indian customer of supply stability for an intermediate' |
| Constraints | Red lines, standards, boundaries? | 'Be serious and professional' | 'Never invent data; cite every figure; no efficacy claims' |
| Format | What should the output look like? | 'Summarize it' | 'One table: dimension / fact / source / confidence' |
| Audience | Who reads it? At what level? | 'Write a report' | 'For the GM: conclusion first, readable in 5 minutes, max 1 page' |
2. Give it all at once, or step by step
Split by nature: stable parts are given once up front; exploratory parts are iterated step by step.
| Type | What goes in | How to give it |
|---|---|---|
| Up front (stable constraints) | Role, compliance red lines, format, audience | Put it at the top and never change it — like a company charter |
| Step by step (exploration) | Outline to detail, sample to batch, one to ten | Confirm each step before the next — pay by milestone |
- 1.Deliver in stages with checkpoints: never demand the whole thing at once. Approve the outline, then the details — a mistake costs one step, not the whole draft.
- 2.End every round by making AI restate your red lines in one sentence. Continue only when it recites them correctly — the cheapest guard against drifting constraints.
- 3.New topic, new chat: mix three tasks in one thread and AI blends them. One task per conversation costs almost nothing.
3. Three drills against hallucination
'Cite the source of every key fact; write NOT FOUND instead of inventing.' Unsourced content is not evidence.
'Have each conclusion rate itself high / medium / low confidence.' Low confidence always goes to the human review queue.
People verify numbers — a wrong figure is worse than none. HS codes, quotes and regulatory conclusions: AI gives leads, not verdicts.
4. Compliance-safe output in pharma
- ·Quote the regulation first, then interpret: paste the original paragraph, keep interpretation in a separate section.
- ·Run external drafts through the three compliance checks: any efficacy claim? any unapproved indication? every figure sourced? One fail = fix before sending.
- ·Mask sensitive data first: customer names, prices and process parameters become 'Customer A / an intermediate / X%'.
- ·AI never replaces regulatory or quality judgment: official text and registration colleagues rule; AI screens and summarizes.
5. The five-step prompt template
Fixed flow: set the role, give context, state constraints, specify format, verify output. Save it and fill the blanks.
'You are a ___' (e.g. a registration consultant fluent in ICH regulations).
'I work at ___ doing ___; the task is ___; the backstory is ___.'
'You must ___; you must not ___' (e.g. no invented data, cite everything, no efficacy claims).
'Output as ___ (table / three sections / one page), about ___ long.'
'Show me ___ (outline / sample) first; continue after I confirm. End by restating all red lines in one sentence.'
6. The three credibility checks
- 1.Where is the source? Verifiable citations raise confidence; polished lines without sources are just leads.
- 2.Is it current? Regulations, prices and registration status change. Ask what date the data is current as of; when unsure, check the official site.
- 3.Can it be verified? If one search or one experiment settles it, ask AI for the verification path. Unverifiable conclusions stay out of decisions.
Unit quiz (instant score + explanations)
Lesson 2 quiz: five questions, instant score with per-question explanations. Nothing is uploaded — results stay in your browser.
Your score stays in this browser only — nothing is uploaded or recorded.
Tools: what to try today
One tool × one real task — prove it, then scale- · Pick tools from three columns: in-house, free to try, on the roadmap
- · Know the data red lines and how to mask
- · Run your first tool-task combo with the two-week pilot
1. The three-column tool map
Pro: Enterprise AI assistant on the company model channel, audio transcription with AI minutes, training-zone AI tutor — onboarding available. Start here
Con: Fixed feature set; niche needs may fall outside
Fit: Everyday office work: writing, calculating, looking up, organizing
Pro: General Q&A, writing, translation, document knowledge bases, light spreadsheet and doc work — good first guinea pigs
Con: Public services — sensitive data must stay out
Fit: Personal trials, calibrating trust, building intuition
Pro: Private knowledge bases (policy and regulation Q&A), agent automation (change digests, briefings), template automation — mapped to several of the 19 questions
Con: Needs pilots first — never start with a big platform
Fit: Teams with a sharp pain point and classified data
2. How to try each column
| Tool | Fits | First step (do it today) | Exercise |
|---|---|---|---|
| In-house AI assistant | Everyday office work | Send today's real request using the four elements | Write one five-step prompt for your own team |
| Transcription + AI minutes | Meetings, customer calls, training recordings | Run a 10-minute recording through transcribe → summarize → action items | Retro-write minutes for one meeting last week |
| General chat tools | Polishing, translation, explaining concepts, brainstorming | Test it with a question whose answer you know — calibrate trust | Ask two tools the same question and compare |
| Document knowledge base | Private doc Q&A (policies, SOPs, regulations) | Load five policy docs, test ten frequent staff questions | Score accuracy; below 80%, improve the documents |
| Template automation | AI draft → company template | Write your template's fonts, colors and layout rules as one constraint block | AI drafts five slides; humans spend five minutes on layout |
3. Data red lines (before anything else)
- ·Mask first: 'Customer A', 'an intermediate', 'X% yield' — the advice stays just as good, the leak risk drops to zero.
- ·Use in-house tools per company rules: internal info stays on the enterprise channel; public tools only see masked or public information.
- ·Human review before anything leaves: AI draft → three compliance checks → human sign-off, for every customer- or regulator-facing document.
4. Trial discipline: the two-week pilot
Each person pilots exactly one combo. No shopping sprees.
What did I ask today, is the conclusion reusable, tag it and file it.
Good? Freeze it into a template and share with one or two peers. Bad? Switch decisively — no loyalty to sunk cost.
Verified prompts enter the team knowledge base so nobody starts from zero again.
5. Private knowledge bases and multi-model setups
- 1.Classify data in three tiers: public (ask anywhere), internal (company tools only), sensitive (mask it or skip). The tier picks the tool, never the reverse.
- 2.Pilot a narrow-topic knowledge base first — never start with a 'company brain'. Pick one small, painful topic (e.g. registration FAQ) and measure accuracy.
- 3.Use off-the-shelf agents first: prove one scenario (read files → organize → output) before building anything yourself.
- 4.Privacy trio: private or enterprise-channel deployment, sensitive-field masking, access logging. Only then discuss multi-model integration.
Unit quiz (instant score + explanations)
Lesson 3 quiz: five questions, instant score with per-question explanations. Nothing is uploaded — results stay in your browser.
Your score stays in this browser only — nothing is uploaded or recorded.
The 19 questions, part I: sales & business (10)
Pain → three-step path → pitfalls: the approach, not the key- · Take away methods and paths, not turnkey solutions
- · Map the patterns onto your own team's scenario
- · Avoid the pitfalls called out in each question
Framing: these questions come from a real corporate AI-needs review and are common across pharma foreign-trade peers. Each follows pain → path → pitfalls; adapt the path to your own systems and data.
Pain: Give all constraints at once or iterate? Fear of AI forgetting rules, contradicting itself, non-compliant output.
- 1.Stable constraints up front, exploration iterated (see Lesson 2)
- 2.Freeze team prompts into the five-step template as reusable assets
- 3.Stage long tasks and make AI restate red lines every round
Pitfall: No magic one-liner exists; a stable process beats inspiration.
Pain: Mixed customer types; real purchasing power judged manually from scattered sources — slow and inconsistent.
- 1.Define a customer-profile field sheet (type, main business, capacity, registrations, purchase history, decision makers, product fit)
- 2.Have AI fill each field from websites, email signatures and trade-show material — every item sourced
- 3.Score into A/B/C tiers with priority advice; feed new emails to AI to refresh the profile — humans set the rating
Pitfall: Unsourced items score zero; AI screens, humans confirm authenticity.
Pain: A dozen stages from inquiry to steady orders; info scattered across email, chat, spreadsheets and personal notes — dropped follow-ups, repeated asks.
- 1.One customer × one product = one file, with fixed fields (stage, milestones, to-dos, owner, open issues)
- 2.Define a stage dictionary (AI picks only from the list); weekly, AI reads the records and outputs stage, to-dos, open issues and risks
- 3.Put key milestones on a calendar; AI whistles early
Pitfall: Keep fields few and fixed — never let AI invent fields. Log verbal agreements the same day; a week later it cannot be reconstructed.
Pain: Strategy by gut feel; no systematic read on market shifts, competitors, customer demand or own competitiveness.
- 1.AI produces one fact card each for market shifts, competitor moves, customer demand and own competitiveness — facts and sources only, no judgment
- 2.Humans combine the four cards into a target-market × customer × product hypothesis
- 3.Test small — three customers for a month; quarterly, AI compares forecast vs. actual
Pitfall: Strategy is human judgment; AI's job is laying out the facts. Never let AI issue the strategic conclusion.
Pain: No systematic read on prospecting volume, conversion, order capability or account care — talent decisions without data.
- 1.Define the four metrics precisely (what counts as a prospect? how is conversion counted?) — definitions before tools
- 2.Get the data into a structured sheet (a spreadsheet is enough to start)
- 3.AI writes a one-page profile per person: strengths, gaps, evidence rows — every claim must cite a data row
Pitfall: No impression-based scoring; delete any profile line without evidence. Growth, staffing and performance decisions stay human.
Pain: The same fact asked repeatedly in one day; changes flow among two or three people; issues scroll away in chat, untraceable later.
- 1.Build one single source of truth (one sheet or board) — enter once, reuse everywhere
- 2.Log every change with who, when, what changed, and who is affected
- 3.Daily, AI extracts changes from chat and email into a digest pushed to the PM and line manager; stale, conflicting or unconfirmed items flag amber; sensitive data gets tiered access
Pitfall: Unify fields and the entry point before automating — automation on chaos only syncs the chaos faster.
Pain: API and intermediate markets swing hard; manual collection lags; price spikes, shortages and supply breaks go unpredicted.
- 1.List what to watch: prices, major producers' capacity, environmental and trade policy, supply-demand structure — 3–5 fixed signal sources per category
- 2.Fixed rhythm: AI digests weekly, reports anomalies anytime; thresholds trigger alerts (price swings, days of inventory)
- 3.Every alert carries a suggested action (lock price / stock up / switch source)
Pitfall: Few and steady: master the three most painful items first. Too many signal sources equals none.
Pain: Market structure, supplier strength and cost structure compared manually and piecemeal — thin dimensions, weak negotiating ground.
- 1.Build the comparison dimension list: market size, capacity spread, process and quality differences, total cost, price history, certifications, substitutes
- 2.AI collects per dimension; certification claims must come from authoritative databases, original text
- 3.Output a comparison matrix with scores, refreshed before negotiations; on new capacity and future supply-demand, AI gives a forecast plus its basis
Pitfall: AI screens only; certifications follow official databases — a wrong certification is worse than none.
Pain: Multiple AIs, inconsistent styles; want a personal-style integrated tool; need privacy and private knowledge base support.
- 1.Classify data three ways (public / internal / sensitive); the tier picks the tool
- 2.Pilot a narrow-topic knowledge base (e.g. registration FAQ) and measure accuracy
- 3.Run one scenario on off-the-shelf agents before building; privacy trio: enterprise channel, masking, access logs
Pitfall: Pilot before platform; integration that cannot save an hour a week is not worth it (Lesson 1).
Pain: Prompts scattered across chats, nothing reviewed or captured, the same questions asked again and again.
- 1.Fix one prompt log (one line a day: question / conclusion / reusable tag)
- 2.Ten minutes before clocking out: AI summarizes and files the day's chats
- 3.Three entries a week into the team knowledge base; build the reflex: search the base before asking
Pitfall: The key is fixed tags and ten minutes a day — not a grand fully-automatic project.
Unit quiz (instant score + explanations)
Lesson 4 quiz: five questions drawn from the ten sales & business questions. Instant score with per-question explanations.
Your score stays in this browser only — nothing is uploaded or recorded.
The 19 questions, part II: R&D (6), HR (2), QA (1)
R&D: hypotheses need proof. HR: definitions and source text. QA: the regulation as written- · Grasp the R&D split: AI hypothesizes, experiments confirm
- · Land the HR pair as an FAQ base and templating
- · Tie all 19 questions together with the three closing rules
1. R&D (6 questions)
Pain: Buyers, labs, manufacturers and literature held as loose contacts — no certifications, no reviews, no alerts, no traceability.
- 1.Unified codes plus required fields (certifications, expiry, reviews, cooperation status, due dates)
- 2.Import the backlog once (AI extracts fields from old sheets)
- 3.Automatic expiry and risk alerts; after each cooperation, force one review entry — the library sharpens with use
Pitfall: Few, solid fields first; add later. Design thirty fields up front and the library never ships.
Pain: Not sure how to state requirements; cannot tell when AI is trustworthy.
- 1.The full method is Lesson 2: the four elements plus the three credibility checks
- 2.This question is the foundation of all 19 — drill it as a team first
- 3.Turn common needs into department templates so nobody starts from zero
Pitfall: Assuming 'AI reads my mind' is the root of all rework — spend three extra minutes stating the need.
Pain: Unknown impurities inferred one experiment at a time — slow and costly; impurity profiles compiled by hand, error-prone.
- 1.Structure the reaction route (substrate, reagents, solvents, conditions, mechanism)
- 2.AI enumerates candidate impurities by mechanism and molecular weight (with formation rationale and confidence) and suggests controls
- 3.Candidates queue for experimental confirmation; template the impurity profile: route → forced degradation → known/potential impurity draft
Pitfall: AI offers hypotheses, not conclusions; impurity structures need experimental proof; control strategy follows regulations and experiments.
Pain: Grading is hard: under-study risks deficiency letters; over-study wastes material and time.
- 1.Lead with three questions: toxic structural alerts? daily dose? process-related or degradant?
- 2.Thresholds and classes cite ICH Q3A(R2)/Q3B(R2)/M7 verbatim (AI finds the text, humans verify)
- 3.AI grades each impurity and suggests study depth; freeze 'enough vs. excessive' into one decision tree for everyone
Pitfall: Thresholds and classes must quote the regulation — AI may not invent them. Once the decision tree is set, changes go through change control.
Pain: Separation, property analysis and method development run on trial and error — many aimless runs.
- 1.AI predicts physicochemical properties (logP, pKa, solubility) to narrow the experiment space
- 2.Design of experiments: AI drafts the run matrix so each run yields more information
- 3.Build a similar-compound method library for AI retrieval; feed every run's result back for review
Pitfall: Predictions are starting points; chromatographic behavior must be measured. AI-suggested conditions get a small trial before scale-up.
Pain: Fragmented, lagging information; risks found too late; forever firefighting.
- 1.One risk register (risk / impact / owner / trigger condition)
- 2.Weekly, AI aggregates cross-department data for warning signs (slipping schedules, change churn, mismatched definitions)
- 3.Triggers alert the PM and line manager automatically; after each handling, backfill the register — alerts sharpen with use
Pitfall: Few and accurate: start with five high-frequency risk signals. Too many false alarms and the alert becomes the boy who cried wolf.
2. HR (2 questions)
Pain: Training reminders and task submissions chased by hand; staff repeatedly ask policy questions — repetitive answering, training fades.
- 1.Build a policy FAQ base first — the 30 most-asked questions with standard answers
- 2.Stand up the assistant on a knowledge-base tool: search first, answer with links to the policy text
- 3.Automate reminders (training and task milestones on a calendar); escalate what it cannot answer, and backfill the FAQ with the human answer
Pitfall: Policy meaning follows the source text; the assistant retrieves and quotes, never improvises — a wrong policy answer is worse than none.
Pain: AI output does not fit the fixed company template; every file hand-reformatted, inconsistent style.
- 1.Turn the template into a rules text: fonts, colors, layout, chart style as one constraint block, given to AI with every prompt
- 2.Fix the flow: AI outline → AI draft → apply template (scripted or manual)
- 3.Build a layout library: five common page types as half-templates, AI fills content only; humans keep a five-minute final review
Pitfall: AI's grip on layout detail is limited — no full automation fantasy. Five minutes of touch-up is already the win.
3. QA (1 question)
Pain: Regulators update often; manual tracking lags; gaps in dossiers, specs and SOPs surface too late.
- 1.Fix the watch list: regulator sites plus email subscriptions / RSS — automate whatever can be automated
- 2.Weekly, AI digests updates and gives a first-pass impact read (which product, which document)
- 3.Build a clause-by-clause checklist against dossiers, specs and SOPs; AI finds gaps and suggests fixes; gaps become tracked remediation tasks
Pitfall: AI does first-pass reads and gap-finding only; interpretation follows official text plus regulatory/QA colleagues, and humans sign compliance conclusions.
4. The three closing rules
- 1.Unify entry points and definitions before automating (#3/#5/#6/#11/#18) — data outside one door leaves AI nothing to organize.
- 2.AI drafts and leads; humans verify and decide (all 19) — especially numbers, certifications and regulations.
- 3.Pilot one scenario, conclude in two weeks, then scale (all 19) — small steps beat grand plans.
Unit quiz (instant score + explanations)
Lesson 5 quiz: five questions from R&D, HR and QA plus the closing rules. Instant score with per-question explanations.
Your score stays in this browser only — nothing is uploaded or recorded.
Workshop and the 30-day plan
One scenario × one tool × 30 days — prove it, then scale- · Finish three drills: rewrite a requirement, pick a scene, set a review card
- · Leave with a one-page 30-day plan
- · Know the closing homework and Q&A routine
1. Drill 1: rewrite with the four elements (15 min)
Take the real requirement your team prepared before class and rewrite it with Lesson 2's five steps: fill in context, set constraints (at least one red line), specify the format, name the audience, and end with 'show the outline first; restate all red lines at the end.'
2. Drill 2: pick your landing scenario (20 min)
'Our team will solve ___' (map to one of the 19 questions, or your own).
Step 1 ___ (this week); step 2 ___ (within two weeks); step 3 ___ (within a month).
What counts as done in 30 days? ___ (quantifiable: hours saved / accuracy rate / documents shipped).
3. Drill 3: set your review card (10 min)
Give your scenario a daily review card, three lines: what did I ask today / is the conclusion reusable (yes·no) / tag (e.g. customer research / regulations / writing). Fill it in ten minutes before clocking out. In two weeks it becomes the seed of your team's knowledge base.
4. The 30-day plan template
| Week | Action | Deliverable | Pass criteria |
|---|---|---|---|
| Week 1 | Pilot: one scenario × one tool, really using it | It is in use, even if clumsy | At least 3 real tasks through it |
| Week 2 | Freeze: good prompts become fixed templates | One prompt template | A colleague can use it too |
| Week 3 | Spread: one or two peers trial your template | Feedback notes | Real feedback received, good or bad |
| Week 4 | Review: reconcile against success criteria, share 10 minutes | One-page review, data-led | Hours saved / accuracy quantifiable |
Close the books by Lesson 1's math: if hours saved exceed hours invested, keep the scenario; if not, switch — no attachment to sunk cost.
5. Closing homework and Q&A routine
- ·Submit the one-page 30-day plan (template above) within one week.
- ·Post a weekly review-card digest in the group (five lines max).
- ·Day 30: ten minutes each on 'one thing I built with AI' — verifiable, real.
- ·Q&A scope: we give approaches — methods, paths, pitfalls. Turnkey solutions follow the path back at your desk. Ask with the four elements — that is practice too.
Unit quiz (instant score + explanations)
Lesson 6 quiz: five questions, instant score with per-question explanations. Nothing is uploaded — results stay in your browser.
Your score stays in this browser only — nothing is uploaded or recorded.
FAQ
Is this a turnkey-solution course?
No. All 19 questions give approaches — methods, paths and pitfalls. Solutions for your own systems are built back at your desk, on your own data.
Are wrong answers recorded? Can anyone else see them?
Nothing is recorded or uploaded. Results live only in your own browser and clear with a refresh or cache wipe — experiment freely.
Where should I start?
In order: Lesson 1 for the math, Lesson 2's four elements (the foundation for all 19), Lesson 3 for tools, then find your team's scenario in Lessons 4–5 and leave with the plan from Lesson 6.
Which tool? Should we pay?
Start with what the company provides, then free options. Depth over breadth: one tool × one real task. Pay only when Lesson 1's math shows at least one person-hour saved per week.