This course remembers completed sections, assessment point totals, and your last course page in this browser on this device. RayAI does not receive this progress or anything you write. Selected answers and written responses are not saved. Clearing browser data, using private browsing, or changing browser or device can remove or hide saved progress.
CAPSTONE | 45 MINUTES COMPLETE / 20 MINUTES FAST TRACK
Build and test a Trust-First prompt workflow
Choose the capstone for your path
Fast Track — about 20 minutes: Use the fictional Source Notes below. Write a safe first prompt and inspect an answer (or diagnose the supplied weak example). If you diagnose a material failure, revise one instruction to address it. If no material failure is found, identify a plausible risk or changed condition and perform a robustness test. Compare the evidence and document your decision; do not invent a failure. Check the $600 limit and Maya’s role against the notes; mark the missing venue/date information. Record your task and owner, safe context, boundaries and output, uncertainty, check, revision and final decision. No invitation is sent and no account is connected.
Complete Course — about 45 minutes: Follow the seven Complete Course steps below with a safe task. Test the initial prompt and both revisions; record the evidence for your check and the remaining limitation. Use the same six-row rubric.
No AI tool available: Use the worked example, identify the invented speaker assignment, write the targeted correction and check it against the Source Notes. Follow the no-tool sequence below for your selected path. Label supplied-output analysis and planned tests separately; do not claim an AI run you performed.
For either path, briefly decide whether a connected service is needed, what access would be excessive, and who would authorize it. Do not connect one for the course. A total score cannot compensate for missing privacy, independent verification or human responsibility: revisit any criterion scored zero.
Before running any prompt, check that every input is non-sensitive and authorized. Never execute the worked example’s unsafe request to send invitations.
Fast Track steps: one revision or robustness test
- Use the fictional Source Notes. Record the task, audience, organizer as decision owner, and what an incorrect answer could cause.
- Write and run a safe first prompt. Keep its output for comparison.
- Inspect the answer. If you diagnose a material failure, change one instruction to address it and run the revised prompt once. If no material failure is found, identify a plausible risk or changed condition, state what you expect the answer to do, and run a purposeful robustness test. For example, add conflicting budget notes and test whether the answer flags the conflict instead of choosing a budget. Document the test; do not invent a failure.
- Compare the two outputs. Record what changed, including no improvement if that is what happened. Independently check the $600 ceiling, Maya's role and unconfirmed details against the notes.
- Use the six-row rubric and record your human decision, remaining limits and whether any connected service is needed. No connection or sending is required. You can now complete Fast Track; the final knowledge check is optional further practice.
No-tool steps for either path
Use this sequence instead of the run instructions. Keep the same safety, source-checking and decision requirements.
- Read the fictional Source Notes and initial prompt. Record the task and human decision owner.
- Inspect the supplied flawed output and identify the unsupported speaker assignment. Write your own targeted prompt revision before reading the supplied correction. Alternatively, if you inspect an output with no material failure, identify a plausible risk or changed condition and plan a robustness test; do not invent a failure.
- Compare the supplied initial and corrected outputs. Check each important claim yourself against the Source Notes. Record this as analysis of supplied outputs, not measured performance of your prompt.
- For Fast Track, one written revision or documented planned robustness test is enough. For Complete Course, write a second targeted revision for a further diagnosed material failure; when none is found, make the second revision test a stated risk or changed condition, such as conflicting budget notes. For each planned test, state the expected result, the source check and comparison you would use, and the decision you would make for each possible result. Compare available supplied evidence and identify what remains untested. Do not invent a failure, output or performed test.
- Score your analysis with the rubric, explain your evidence and final decision, and name what remains untested. This demonstrates practice with a supplied example; it does not demonstrate unaided performance on a new task.
Complete Course steps: two revisions with an AI tool
These seven steps apply to Complete Course learners using an AI tool. Fast Track learners use the shorter sequence above; learners without a tool use the no-tool sequence.
Choose the task. Write what you want to accomplish, who will use the answer, who must make the final decision, and what could go wrong if the answer is incorrect.
Run your first prompt. Open your AI assistant, enter your original prompt, send it, and save the answer.
Write an improved prompt. Make the task, needed facts, and output clear. Add the audience, limits, purpose, and checking instructions that your task needs. Tell the AI what to do when information is missing.
Run the improved prompt. Compare the new answer with the first answer. Record what changed (including no improvement) and one remaining problem or test condition.
Check one important fact, number, source, promise, or missing detail using an authoritative source, reliable calculation, or qualified person.
Make the second revision. If you diagnose a further material failure, change only the instruction needed to address it and run the revised prompt. If no further material failure is found, state a plausible risk or changed condition, such as conflicting source notes, and revise the instruction for handling it. Run this robustness test and compare the evidence with the earlier answer and source. Record the result and remaining limits; do not invent a failure.
Decide what to do. Choose: Use it / Improve it again / Ask a qualified person / Do not use it. Explain your choice.
Completed example: from error to checked draft
Fictional teaching example. Maya is an invented character. Both outputs below are written examples, not records of a live AI run. No private data, messages or account connections are involved. You can complete this comparison on paper without an AI tool.
1. Source notes. Community Skills Night is a practical evening for adult community members to share useful skills. Budget ceiling: $600. Maya offered to contact two possible speakers. A venue request was submitted but not confirmed. Date, capacity, accessibility, agenda and participant consent process are not confirmed.
2. Initial prompt. Use only these Source Notes to write a short planning brief for volunteer organizers. The organizer owns the decision about outreach. Include purpose, budget, speaker roles, open questions and next actions. Mark unsupported details "Needs confirmation." Do not send, publish or contact anyone. [Append the Source Notes above.]
3. Flawed initial output. "Purpose: adults share useful skills. Budget ceiling: $600. Speaker: Maya. Venue, date, capacity, accessibility, agenda and consent process: Needs confirmation. Proposed next action: organizer checks the missing details before outreach."
4. Material error. "Speaker: Maya" turns an offer to contact possible speakers into a commitment to speak. Invitations based on this claim would misrepresent her role. The source supports contact only, and confirms no speaker. A source-only instruction did not prevent the error.
5. Targeted revision. Keep the initial prompt and notes; add: "Maya offered only to contact two possible speakers. Do not identify her as a speaker. Separate the confirmed contact role from unconfirmed speaker assignments."
6. Corrected output. "Purpose: adults share useful skills. Budget ceiling: $600; spending is not established. Confirmed role: Maya offered to contact two possible speakers. Speaker assignments: Needs confirmation. Venue request: submitted, not confirmed. Date, capacity, accessibility, agenda and consent process: Needs confirmation. Proposed next action: organizer checks these gaps before outreach. No invitation is sent."
7. Human verification. I compared both outputs directly with the Source Notes: $600 is a ceiling, not an expense; Maya's contact role is supported, a speaker assignment is not; a submitted venue request is not a booking. The corrected output preserves each distinction. Asking the model to review itself would not replace this check. In a real task, the organizer must separately confirm venue capacity and accessibility through an official channel before outreach; that external check has not occurred in this fictional example.
8. Before and after. Before: "Speaker: Maya" - unsupported commitment. After: "Maya offered to contact two possible speakers; speaker assignments need confirmation" - traceable to the notes. The revision fixes the observed role error; it does not establish the missing event facts.
9. Rubric self-evaluation. Purpose and owner: 2 (organizer, audience and outreach decision explicit). Prompt design: 2 (task, format and boundaries explicit). Evidence: 2 (source-only rule and gaps traceable). Privacy and safety: 2 (fictional minimum data; no connection or sending permission needed). Verification: 2 (human checks the material role and budget claims against the supplied source). Iteration and decision: 2 (specific failure, targeted revision, comparison and decision). Total: 12/12 for this supplied example, not evidence of your own performance or a credential. A score of 2 for source verification does not mean the venue was externally confirmed.
Decision. Keep the corrected brief as an internal draft. Do not draft or send invitations until the organizer confirms the missing venue, date, accessibility, consent and recipient-authority details. For your own record, explain your checks and any lower scores. With no AI tool, label your work "analysis of supplied outputs and planned test"; do not claim an AI run.
Optional harder test: conflicting budget notes
Add an undated note saying "Budget $800" while the original says "Budget ceiling $600." A successful response flags the conflict, avoids choosing either as the current authorized budget, and asks the organizer for the authoritative approval. Revise, compare and record that unresolved limit; do not average the figures.
Capstone rubric
| Criterion | 0 - Missing | 1 - Partial | 2 - Strong |
|---|---|---|---|
| Purpose and owner | No real outcome or owner | Outcome or owner is vague | Outcome, audience, decision, owner, and consequence are clear |
| Prompt design | Task is vague | Some useful elements | Task, context, output, audience, boundaries, purpose, and quality fit the risk |
| Evidence | No source rule | Sources named but gaps unclear | Evidence boundary, missing-information behavior, and source traceability are explicit |
| Privacy and safety | Sensitive/action risks ignored | Some caution but incomplete | Data is necessary, non-sensitive and authorized; connection permissions and action controls are justified |
| Verification | No check | Self-check only | At least one material item is checked independently with a suitable method |
| Iteration and decision | No targeted revision or purposeful robustness test, or no human decision | Revision or robustness test is attempted, but its reason, evidence comparison or human decision is unclear | A diagnosed failure drives a targeted revision; or, when no material failure is found, a clearly stated risk or changed condition drives a purposeful robustness test. The learner compares the evidence and makes a responsible decision. |
Your capstone record
Label your record "AI runs" or "supplied-example analysis and planned tests". Keep original prompts and outputs in your own non-sensitive notes. Complete Course: include both revisions and their observed results or clearly labeled planned tests. These fields are not saved.
Finish this capstone
Mark this capstone complete when you have attempted every activity you intend to complete and reviewed your work against the course guidance.