Twelve chapters on prompt design, document analysis, task-based assistant comparison and data safeguards. Synthetic training cases and data; not real ISHRAQA7 performance.
Twelve-chapter curriculum
- AI assistants and their roles
- Getting started with Gemini and Claude
- Building comparable prompts
- Iterating and rating responses
- Drafting and document synthesis
- Research, evidence and verification
- Analysing files without invention
- Creating professional deliverables
- Projects and collaboration
- Comparing assistants without arbitrary rankings
- Assistant security and confidentiality
- Capstone: comparative document workflow
1. AI assistants and their roles
A fictional training centre must summarise ten requests without automatically contacting any customer.
Method and workflow — TRACE method
An assistant can accelerate drafting and organise supplied material; it is neither a documentary authority nor a decision-maker. Separate generated suggestions, accessible evidence and the owner’s approval of the file. The outcome of this chapter is observable: Scope-and-approval table. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Choose a low-risk task, specify available input and define an observable success rule. Keep a log of input, evidence, output, discrepancy and human decision. Repeat with the same dataset in both tools if accounts are available. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: A fictional training centre must summarise ten requests without automatically contacting any customer. Deliver scope-and-approval table. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Confident wording is not evidence; no draft may be sent without review.
Knowledge check
- Inputs can be omitted.
- Deliver: Scope-and-approval table.
- An unverified answer is official policy.
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Follow the method: Choose a low-risk task, specify available input and define an observable success rule. Keep a log of
- Apply this safeguard: Confident wording is not evidence; no draft may be sent without review.
- Run without authorisation.
- Publish the first output immediately.
2. Getting started with Gemini and Claude
A team writes an onboarding card for colleagues with different account types.
Method and workflow — TRACE method
Interfaces and entitlements vary by plan, region and account. Training should identify core interactions without assuming that every button or advanced feature is available to everyone. The outcome of this chapter is observable: Dated availability sheet: confirmed / limited / unverified. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Open both official services when authorised. Record date, access tier, visible interface, attachment options and disclosed limits. Otherwise use documented, dated examples clearly labelled as illustrative. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: A team writes an onboarding card for colleagues with different account types. Deliver dated availability sheet: confirmed / limited / unverified. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Do not invent menus or prices; consult official help before writing click-by-click instructions.
Knowledge check
- An unverified answer is official policy.
- Inputs can be omitted.
- Deliver: Dated availability sheet: confirmed / limited / unverified.
- Follow the method: Open both official services when authorised. Record date, access tier, visible interface, attachment
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Publish the first output immediately.
- Apply this safeguard: Do not invent menus or prices; consult official help before writing click-by-click instructions.
- Run without authorisation.
3. Building comparable prompts
Two assistants receive five fictional meeting notes and must produce a decision table.
Method and workflow — TRACE method
A fair comparison requires the same context, input dataset, scoring criteria and output schema. Wording may adapt to interfaces, but the purpose and constraints must remain fixed. The outcome of this chapter is observable: Master prompt and two anonymised run logs. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Use a common brief with goal, supplied evidence, constraints and output format. Require an explicit missing-information label. Archive the exact input sent to each assistant for reproducibility. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: Two assistants receive five fictional meeting notes and must produce a decision table. Deliver master prompt and two anonymised run logs. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Changing inputs across tools invalidates the comparison.
Knowledge check
- Deliver: Master prompt and two anonymised run logs.
- An unverified answer is official policy.
- Inputs can be omitted.
- Ignore missing cases for speed.
- Follow the method: Use a common brief with goal, supplied evidence, constraints and output format. Require an explicit
- Change evidence during a comparison.
- Run without authorisation.
- Publish the first output immediately.
- Apply this safeguard: Changing inputs across tools invalidates the comparison.
4. Iterating and rating responses
An initial summary misses two actions and invents a delivery date.
Method and workflow — TRACE method
The first output is only a working hypothesis. A protocol improves accuracy, coverage, clarity and caution by changing one condition at a time. Polished but fabricated content fails the quality gate. The outcome of this chapter is observable: V0/V1 rubric with corrections and justified acceptance. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Prepare five fixed tests, a four-criterion rubric and a pass threshold. Compare V0 and V1, annotate evidence and decide whether the improvement comes from a controlled change rather than chance. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: An initial summary misses two actions and invents a delivery date. Deliver v0/v1 rubric with corrections and justified acceptance. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
A model’s self-score never substitutes for independent checking.
Knowledge check
- Inputs can be omitted.
- Deliver: V0/V1 rubric with corrections and justified acceptance.
- An unverified answer is official policy.
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Follow the method: Prepare five fixed tests, a four-criterion rubric and a pass threshold. Compare V0 and V1, annotate
- Apply this safeguard: A model’s self-score never substitutes for independent checking.
- Run without authorisation.
- Publish the first output immediately.
5. Drafting and document synthesis
A fictional committee approves a course but assigns neither an owner nor a budget.
Method and workflow — TRACE method
A strong summary preserves approved decisions, cites its sources and flags missing information. Gemini and Claude may suggest different structures; choose based on audience needs rather than product preference. The outcome of this chapter is observable: 250-word minutes plus a missing-information register. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Prepare two reference documents, an expected outline and a meeting-minutes template. Ask for an outline before the draft. Trace every factual sentence to numbered source material and list unresolved points separately. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: A fictional committee approves a course but assigns neither an owner nor a budget. Deliver 250-word minutes plus a missing-information register. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Never turn an assistant suggestion into an approved decision.
Knowledge check
- An unverified answer is official policy.
- Inputs can be omitted.
- Deliver: 250-word minutes plus a missing-information register.
- Follow the method: Prepare two reference documents, an expected outline and a meeting-minutes template. Ask for an outl
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Publish the first output immediately.
- Apply this safeguard: Never turn an assistant suggestion into an approved decision.
- Run without authorisation.
6. Research, evidence and verification
Two outputs claim a service has a new feature but provide no release note.
Method and workflow — TRACE method
A factual claim needs an accessible dated source even when several assistants repeat it. Two similar model answers are not independent evidence. Separate supporting documents, interpretation and uncertainty. The outcome of this chapter is observable: Evidence ledger with a status for each claim. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Build a five-claim evidence register with official URL, access date, relevant passage, reasoned confidence and limitations. Mark a claim unverified if you cannot locate primary evidence. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: Two outputs claim a service has a new feature but provide no release note. Deliver evidence ledger with a status for each claim. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Never accept a fabricated citation as a reference.
Knowledge check
- Deliver: Evidence ledger with a status for each claim.
- An unverified answer is official policy.
- Inputs can be omitted.
- Ignore missing cases for speed.
- Follow the method: Build a five-claim evidence register with official URL, access date, relevant passage, reasoned conf
- Change evidence during a comparison.
- Run without authorisation.
- Publish the first output immediately.
- Apply this safeguard: Never accept a fabricated citation as a reference.
7. Analysing files without invention
Eight training expenses contain one duplicate and one out-of-period amount.
Method and workflow — TRACE method
An attachment may be incomplete, badly encoded or sensitive. Analyse its fields, units, period and missingness before making any claim or chart. Document each transformation as an auditable step. The outcome of this chapter is observable: Cleaned table, exclusion log and independently calculated total. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
On a training CSV, check row count, duplicate IDs, totals and dates. Request a table summary, then recompute key figures independently and explain any discrepancy. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: Eight training expenses contain one duplicate and one out-of-period amount. Deliver cleaned table, exclusion log and independently calculated total. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Never upload personal data without authorisation and appropriate safeguards.
Knowledge check
- Inputs can be omitted.
- Deliver: Cleaned table, exclusion log and independently calculated total.
- An unverified answer is official policy.
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Follow the method: On a training CSV, check row count, duplicate IDs, totals and dates. Request a table summary, then r
- Apply this safeguard: Never upload personal data without authorisation and appropriate safeguards.
- Run without authorisation.
- Publish the first output immediately.
8. Creating professional deliverables
A small firm needs a bilingual memo describing an internal procedure.
Method and workflow — TRACE method
Document quality depends on readability, audience fit, traceability and accuracy. An assistant can propose layouts and tables; the author must verify names, branding, sources and image rights. The outcome of this chapter is observable: Structured memo with accessibility and approval checklist. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Write a brief naming reader, length, tone and format. Compare two layout plans, then apply accessibility checks for contrast, headings, alt text and document language. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: A small firm needs a bilingual memo describing an internal procedure. Deliver structured memo with accessibility and approval checklist. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Never publish unlicensed marks or simulated screenshots presented as authentic.
Knowledge check
- An unverified answer is official policy.
- Inputs can be omitted.
- Deliver: Structured memo with accessibility and approval checklist.
- Follow the method: Write a brief naming reader, length, tone and format. Compare two layout plans, then apply accessibi
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Publish the first output immediately.
- Apply this safeguard: Never publish unlicensed marks or simulated screenshots presented as authentic.
- Run without authorisation.
9. Projects and collaboration
A fictional team drafts one onboarding guide using both assistants.
Method and workflow — TRACE method
A useful AI project defines roles, data, approval and schedule. A saved conversation is not an operating procedure until it becomes a reproducible template and decision record. The outcome of this chapter is observable: Project charter, ownership matrix and version log. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Create a project charter with owner, goal, approved sources, indicators and review process. Simulate two stakeholder comments and log changes without exposing unapproved material. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: A fictional team drafts one onboarding guide using both assistants. Deliver project charter, ownership matrix and version log. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Sharing permissions depend on the account; never assume access controls.
Knowledge check
- Deliver: Project charter, ownership matrix and version log.
- An unverified answer is official policy.
- Inputs can be omitted.
- Ignore missing cases for speed.
- Follow the method: Create a project charter with owner, goal, approved sources, indicators and review process. Simulate
- Change evidence during a comparison.
- Run without authorisation.
- Publish the first output immediately.
- Apply this safeguard: Sharing permissions depend on the account; never assume access controls.
10. Comparing assistants without arbitrary rankings
A marketing memo declares one tool always better after a single trial.
Method and workflow — TRACE method
A defensible comparison measures a task, sample and date, not universal superiority. Use the same rubric for fidelity, completeness, editing effort and risk, and expose the limits of each observation. The outcome of this chapter is observable: Documented task-specific matrix with no overall winner. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Keep five anonymised trials; score each tool on four 0–4 criteria using supporting traces. If access is unavailable, mark the tool not tested instead of guessing a score. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: A marketing memo declares one tool always better after a single trial. Deliver documented task-specific matrix with no overall winner. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Pricing and availability change; cite a dated official source.
Knowledge check
- Inputs can be omitted.
- Deliver: Documented task-specific matrix with no overall winner.
- An unverified answer is official policy.
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Follow the method: Keep five anonymised trials; score each tool on four 0–4 criteria using supporting traces. If access
- Apply this safeguard: Pricing and availability change; cite a dated official source.
- Run without authorisation.
- Publish the first output immediately.
11. Assistant security and confidentiality
A manager wants to summarise a real employee complaint in an unapproved account.
Method and workflow — TRACE method
Material sent to an external service needs classification first. Responsible use separates public, internal, confidential and personal information and follows both account controls and organisational rules. The outcome of this chapter is observable: Refusal protocol and synthetic-data template. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Create a sensitivity matrix. Replace names, emails and identifiers with synthetic values. Define stop conditions for HR and financial material and document human approval and retention. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: A manager wants to summarise a real employee complaint in an unapproved account. Deliver refusal protocol and synthetic-data template. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
Do not treat privacy terms as identical across all plans.
Knowledge check
- An unverified answer is official policy.
- Inputs can be omitted.
- Deliver: Refusal protocol and synthetic-data template.
- Follow the method: Create a sensitivity matrix. Replace names, emails and identifiers with synthetic values. Define sto
- Change evidence during a comparison.
- Ignore missing cases for speed.
- Publish the first output immediately.
- Apply this safeguard: Do not treat privacy terms as identical across all plans.
- Run without authorisation.
12. Capstone: comparative document workflow
A fictional leadership team needs a verified decision memo without ranking providers.
Method and workflow — TRACE method
The capstone integrates prompt design, fair comparison, evidence review, confidentiality and human sign-off. Its output is not a pasted model answer but a reproducible auditable procedure. The outcome of this chapter is observable: Final dossier: prompts, evidence, two drafts, comparison, audit log and sign-off. A professional result keeps evidence, assumptions and decision owners visible. Any assertion without accessible proof remains pending rather than being inferred from polished language.
Prepare two fictional sources, ten requests and a common scoring rubric. Produce or simulate two drafts, check each figure, record differences and issue a human-approved report. Mark any unperformed live-service trial not tested. Check scope, input quality, transformations, output format and approval in that order. Preserve the first attempt and justify each edit. Never hide unresolved cases simply to make the presentation look complete.
Example prompt
Independent exercise
Work through this scenario: A fictional leadership team needs a verified decision memo without ranking providers. Deliver final dossier: prompts, evidence, two drafts, comparison, audit log and sign-off. Create two versions using the same source data, check consistency, log discrepancies and name the required approving role.
If either service is unavailable, label the simulation; do not claim a live trial.
Knowledge check
- Deliver: Final dossier: prompts, evidence, two drafts, comparison, audit log and sign-off.
- An unverified answer is official policy.
- Inputs can be omitted.
- Ignore missing cases for speed.
- Follow the method: Prepare two fictional sources, ten requests and a common scoring rubric. Produce or simulate two dra
- Change evidence during a comparison.
- Run without authorisation.
- Publish the first output immediately.
- Apply this safeguard: If either service is unavailable, label the simulation; do not claim a live trial.
