A hands-on, structured, verifiable twelve-chapter course. Fictional training data, not real ISHRAQA7 performance.
Twelve-chapter curriculum
- Generative AI: principles, capabilities and limits
- Using ChatGPT and organizing your work
- Writing precise, verifiable prompts
- Improving outputs through iteration and evaluation
- Professional writing, synthesis and documents
- Research, sources and fact-checking
- Analyzing files, tables and data
- Creating responsible editorial and visual content
- Building reusable productivity workflows
- Advanced capabilities and durable methods
- Privacy, bias, rights and human oversight
- Capstone: a complete, verifiable document workflow
1. Generative AI: principles, capabilities and limits
A fictional service company receives eighteen customer requests. It only wants to sort them into information, support and complaint. An assistant can propose triage labels; it cannot know the contract or promise a response. An ambiguous request must remain marked as needing confirmation.
Method and workflow — CLEAR method
Generated output is neither evidence nor access to an infallible database. A conversational model creates plausible language from the supplied context and patterns learned in training. It can help organize ideas, draft text and explore alternatives, yet a polished answer may contain a wrong date, an invented citation or an unchecked calculation. Distinguish drafting, analysis of supplied material and external research: these tasks require different evidence standards.
Before prompting, write a one-sentence objective, define which data may be shared and assign a human reviewer. For drafting, check clarity and fit. For analysis, reconcile figures against source rows. For research, open the cited documents and examine author, date and scope. Label important claims as verified, plausible or requiring validation.
Example prompt
Independent exercise
Open ia01-demandes-clients-en.csv, classify at least ten records and identify three decisions requiring human confirmation. Re-check every suggested category against the original sentence.
Fluent language is not proof. Never upload real personal customer information for this exercise.
Knowledge check
- Every citation displayed by AI must be real.
- Fluency does not establish factual accuracy.
- Long answers guarantee accuracy.
- Ambiguous cases must be classified randomly.
- All ambiguous rows should be deleted.
- Ambiguous customer cases should be marked for confirmation.
- Reconcile generated numbers with source data.
- AI numbers never require checking.
- Charts alone validate all calculations.
2. Using ChatGPT and organizing your work
A service manager prepares reply guidance for colleagues. Three deliverables are required: an intake register, response drafts and an approval checklist. Combining them in an unstructured request often hides missing assumptions.
Method and workflow — CLEAR method
Assistant interfaces evolve by plan, device and workspace configuration. The durable skill is organizing a brief, limiting context and documenting decisions, not memorizing button locations. One sprawling conversation may mix unrelated goals and make decisions hard to trace. Separate work by deliverable and keep verified decisions in an external project log.
Create a fictional working folder named Customer welcome pilot. Write five lines covering outcome, audience, permitted data, format and reviewer. Use one conversation for request classification and a separate one for drafting. At each milestone ask for a decision summary; record only checked statements in the project log.
Example prompt
Independent exercise
Write a 120-word project brief and a decision log recording date, assumption, evidence and approving person. Revise the brief when you discover that priority data is missing.
Check which organization features your own account actually provides; menus vary by plan and interface.
Knowledge check
- Starting without a goal improves traceability.
- Every task should share one long conversation.
- Define a deliverable before opening a conversation.
- Separate deliverables to simplify review.
- A single chat replaces a decision log.
- Assumptions should never be recorded.
- Settings never change over time.
- Check actual interface availability.
- Every account has identical menus.
3. Writing precise, verifiable prompts
A team needs a brief note based on eighteen fictional requests. Its manager has two minutes to read it. The note must state trends and uncertainties, without invented quantities or resolution dates.
Method and workflow — CLEAR method
A useful prompt is not a magic phrase. It turns a task into an output contract covering outcome, audience, sources, constraints and success criteria. 'Make a good summary' specifies none of these. 'Summarize for a nontechnical team in 150 words and separate verified facts from open questions' produces an output that can be checked.
Use the CLEAR structure: Context, Learning goal, Evidence, Allowed limits and Requested format. State nonnegotiable items, then instruct the assistant to label missing information instead of guessing. Distinguish task instructions from quoted source material. Review the output against the checklist yourself.
Example prompt
Independent exercise
Rewrite three vague instructions ('summarize', 'make a plan', 'improve the text') as complete CLEAR prompts. Create a five-criterion rubric and score two example outputs from zero to two per criterion.
Asking an assistant to 'verify everything' is not independent verification.
Knowledge check
- A prompt includes explicit success criteria.
- Good prompts never need context.
- Output length is never relevant.
- Vague prompts are reliably more accurate.
- Keep supplied evidence separate from instructions.
- Remove all format constraints.
- Missing fields justify plausible inventions.
- Unknown names can be selected at random.
- Flag missing information rather than guessing.
4. Improving outputs through iteration and evaluation
An early triage prompt forces all tickets into three categories. Review shows two ambiguous entries wrongly labeled. A later version adds needs confirmation and an evidence field.
Method and workflow — CLEAR method
Prompt improvement should be a controlled cycle, not a random sequence of rewordings. Fix a test set, keep input data stable and change one factor at a time: instructions, examples, format or evaluation criterion. Scoring must separate accuracy, completeness, clarity and risk. Without stable test cases, a new response may merely look better because the question changed.
Create three repeatable test tasks: classification, synthesis and drafting. Score the first result on four criteria from zero to four. Change one instruction and repeat with the same input. Keep both outputs and accept the revision only when it improves quality without breaching a critical constraint.
Example prompt
Independent exercise
Run or simulate V0, V1 and V2 on five requests. Log the single variable changed, resulting score and a justified stopping rule.
Model-generated self-scores may be biased; spot-check ratings against the supplied evidence.
Knowledge check
- Testing needs no success criteria.
- Compare on a stable set of test cases.
- Changing everything at once improves attribution.
- Longer output is always superior.
- Appearance outranks factual quality.
- Change one factor at a time when evaluating.
- Reject critical invented facts even if a score rises.
- Ignore errors when scores increase.
- Publish the latest result without review.
5. Professional writing, synthesis and documents
A fictional training center prepares meeting minutes: three decisions were made, but no delivery date was agreed. The minutes must show each action and owner while keeping the deadline as not set.
Method and workflow — CLEAR method
AI-assisted writing should support editorial judgment rather than replace it. An effective meeting record separates facts, decisions and actions; an email states audience and requested action without inventing deadlines; a strategic note distinguishes assumptions. Elegant tone does not cure missing evidence.
Provide a writing brief naming audience, channel, tone, approved facts and structure. Use two stages: outline and missing-information check, then draft. Mark uncertain statements. Review people, dates, obligations and any wording that might create a commitment before sharing.
Example prompt
Independent exercise
Produce minutes and a short colleague email from fictional notes. Compare the two documents for factual fidelity, level of detail and ambiguity.
Never invent quotations from attendees or contractual commitments absent from the source notes.
Knowledge check
- Mix assumptions with approved decisions.
- Inventing deadlines makes text professional.
- Separate facts, decisions and actions in minutes.
- Use approved information when drafting.
- Fluency licenses invented quotations.
- Tone makes name checking unnecessary.
- Delete actions without known owners.
- Mark a missing deadline as not set.
- Pick a random date.
6. Research, sources and fact-checking
An internal brief says an AI service includes a particular function. A manager must check whether their own workspace can use it, without treating an old blog post or advertising claim as authoritative evidence.
Method and workflow — CLEAR method
Assisted research has two parts: locating evidence and interpreting it correctly. A URL alone is insufficient; you must open the document and check the specific claim. Prices, models and legal requirements change over time. If search is unavailable, an assistant should state that current availability cannot be confirmed rather than implying it was checked.
Create a claim register with claim, primary source, checking date and verification status. For each important statement, consult official documentation, capture the relevant passage and record limitations. Distinguish missing source, inconclusive source and verified claim. Add a date to every time-sensitive statement.
Example prompt
Independent exercise
Take three claims from the fictional register. Identify primary sources and complete the verification statuses. Where evidence is missing, write a transparent conditional statement.
'The assistant cited a source' is different from 'I read and checked the source'.
Knowledge check
- Open and check the cited source.
- Any link confirms a claim regardless of content.
- Current information needs no date.
- Prices never change.
- Use primary documentation for availability claims.
- Advertising is definitive technical evidence.
- Treat uncertainty as fact.
- Avoid keeping verification dates.
- Label unverified information clearly.
7. Analyzing files, tables and data
A fictional team compares six training campaigns. It needs total spending and a list of budget overruns. One campaign is on hold, so its inclusion in the total must be defined explicitly.
Method and workflow — CLEAR method
Useful analysis starts with a data dictionary: every column has a meaning, unit and handling rule. Tables may contain duplicates, missing values or incompatible units. Before reporting a KPI, inspect dates, scope, row counts and calculation rules. Independently reproduce important calculations; a polished chart is not proof of correctness.
Open ia01-budget-campagne-en.csv, document its fields and units (MAD), and check time periods. Recalculate totals from rows rather than accepting a single AI-generated figure. Separate budgeted and actual spending, and define variance as actual minus planned. Verify at least two records manually.
Example prompt
Independent exercise
Recalculate planned and actual spending for three campaigns, compare with model output and explain each difference using source records or a stated exclusion rule.
Never present a total without documenting its status filter, dates and currency.
Knowledge check
- Ignore duplicate records by default.
- Check units and scope before computing KPIs.
- Add incompatible currencies directly.
- Accept totals without a formula.
- A tidy chart verifies all arithmetic.
- Recompute totals from source rows.
- Document inclusion and exclusion rules.
- Delete statuses silently.
- Treat a chart as the sole evidence.
8. Creating responsible editorial and visual content
ISHRAQA7 Académie drafts a fictional workshop announcement about checking AI-generated claims. The brief provides no price or confirmed date. The content must remain attractive without inventing event details or testimonials.
Method and workflow — CLEAR method
AI-assisted content accelerates ideation, but editorial quality requires a strategy: audience, purpose, message, channel, visual identity and source checking. Distinguish documentary photography, an illustrative generated scene and a genuine product screenshot. A generated image is not evidence of a real event or actual interface. Review image rights, trademarks and personal likenesses before publication.
Write a creative brief covering promise, angle, audience, forbidden claims, CTA and evidence sources. Draft three accurate headline candidates rather than optimizing only for clicks. For images specify the scene, setting, light, framing, alt text and illustrative status. Seek editorial approval for the combined package.
Example prompt
Independent exercise
Create a magazine card, social caption and longer description. Provide accurate image alt text and a prepublication claim-check list.
Do not fabricate endorsements, statistics, official logos or deceptive screenshots.
Knowledge check
- Present generated imagery as evidence of an event.
- Invent testimonials to build trust.
- Distinguish genuine photography from illustration.
- Stay faithful to audience and brief.
- Click-through rates justify sensational misinformation.
- Image rights never matter.
- Hide all information about image provenance.
- Write descriptive alt text.
- Use keyword stuffing as alt text.
9. Building reusable productivity workflows
A team prepares a Friday request summary. If source records are missing, the workflow must stop. No message is sent automatically to any recipient.
Method and workflow — CLEAR method
A professional workflow defines who provides inputs, which transformations are authorized and who approves the output. Reusable prompts need named variables, exception handling and stopping rules. Automating an unclear decision does not make it reliable. Responsibility matters when content could reach customers or involve confidential documents.
Map intake, input checks, preparation, assisted drafting, independent checking, human approval and archiving. State success and stop criteria for every step. Build a prompt template with explicit variables. Test once with valid fictional tickets and again with a missing mandatory field.
Example prompt
Independent exercise
Draw a simple workflow and draft a five-variable prompt template: period, audience, source, format and reviewer. Test normal data and data without a reporting period.
A documented workflow may remain manual; do not call it automated when no automation is implemented.
Knowledge check
- Define stop conditions before distribution.
- Automatically send every draft.
- Remove human approval.
- Invent missing input values.
- Test missing mandatory fields.
- A workflow must always continue.
- Document only the first screen.
- Ignore decision logging.
- Document input, output and owner for each stage.
10. Advanced capabilities and durable methods
A team wants a chart from the training budget. Some accounts may provide a data-analysis tool while others may not. The required outcome remains a recomputed metric and clearly labeled chart.
Method and workflow — CLEAR method
Assistant capabilities evolve with product, plan and geography: file analysis, web search, workspace tools, image creation and service connections. A responsible course never promises a button to every learner. Teach prerequisite checking—account access, authorization, supported files and limits—and provide an alternative when a feature is unavailable.
Build a matrix of need, capability, observed availability, restriction and fallback. Consult dated official guidance. Run a short non-sensitive test and record whether a feature was actually used or merely described. Prepare a manual option, such as exporting a CSV and verifying calculations independently.
Example prompt
Independent exercise
Create an access matrix for research, files, images and collaboration. Record only what is observed in your own account, and specify a fallback for unavailable capabilities.
Avoid undated claims about prices, quotas or model versions.
Knowledge check
- Product features never change.
- Check capabilities for the actual account.
- Every account has identical access.
- Stop learning if an optional tool is missing.
- Assume premium access without evidence.
- Provide fallbacks for unavailable functions.
- Record tests actually performed.
- Claim a feature was tested when only described.
- Quote quotas without a verification date.
11. Privacy, bias, rights and human oversight
A fictional HR team wants to assess applications. An assistant might help draft objective job-related criteria but must not autonomously decide employment outcomes or infer sensitive attributes.
Method and workflow — CLEAR method
Risk depends on context, not only on the model. Fictional requests can be used for practice; real case files with names, contracts or health data require a separate privacy and authorization assessment. Assistants may reproduce bias, and generated content may raise copyright or trademark questions. Put human oversight before consequential action.
Classify information as public, internal, confidential or prohibited. Document purpose, retention and permitted access according to organizational policy. Replace personal identifiers with practice codes. Review advice from multiple perspectives, record exceptions and establish a written approval step.
Example prompt
Independent exercise
Review ten example data fields (name, phone, skill, API secret, etc.), decide what belongs in a practice dataset and prepare a risk register plus approval checkpoint.
Retention and privacy settings depend on service and workspace: consult the relevant current documentation.
Knowledge check
- Upload passwords to speed up work.
- HR files need no special care.
- Remove unnecessary sensitive information.
- Document risk and human approval.
- AI alone should determine employment outcomes.
- Assistant recommendations cannot be biased.
- Generated content never raises third-party rights issues.
- Review rights and image provenance.
- Brand marks have no usage rules.
12. Capstone: a complete, verifiable document workflow
A fictional training organization wants the distribution of incoming requests and spending across six campaigns. It is not requesting predictions or autonomous commercial decisions. A sourced report with assumptions and proposed actions is sufficient.
Method and workflow — CLEAR method
The capstone combines briefing, practice-data analysis, drafting and quality assurance. It can be completed manually without a paid account; if an assistant is available, its use is optional and documented. Quality is not word count: it is a traceable chain in which every number links back to source rows and uncertainty is explicit.
Create a Monthly service report folder. Define the audience and acceptance criteria; check the eighteen practice requests and fictional campaign budget; write a 250-word factual summary; draft an unsent circulation email; assemble a claim register and unresolved-issues list; request human peer approval.
Example prompt
Independent exercise
Submit four deliverables: report, claims register, calculation checks and an unsent transmission email. Explain two model limitations and one decision postponed for lack of evidence.
Generated output is not an approved project until a responsible person has reviewed its evidence and risks.
Knowledge check
- Link every reported number to source rows.
- Invent metrics for missing data.
- Present assumptions as established facts.
- Ignore ambiguous records.
- Keep a verification register.
- Discard evidence after generation.
- Automatically send generated emails.
- Trust only the assistant's self-score.
- Get human approval before distribution.
