Evaluate, Document and Keep Learning: Your Charter and Next Steps

You now have a library of prompts that seem to work. This final lesson turns "seems to work" into "I know how well it works, and where it fails." You will build a simple evaluation grid and a set of tests, run a few improvement rounds, write a personal charter for responsible use, and plan how to keep learning in a field that changes quickly.

This is the culmination of the course: it combines the testing method from chapter 3, the reliability and legal awareness of chapter 4 and the workflow of chapter 5. Take it step by step and be honest with yourself: finding a weakness is a success, not a failure.

Step 1: Build the evaluation grid

Take the success criteria from your scope document and turn them into a grid. For each prompt and each test case, you will give a simple score. A three-level scale is enough and is faster than a ten-point scale:

• 2 = meets the criterion (ready to use);
• 1 = partly meets it (needs a small fix);
• 0 = fails it (wrong, missing or unsafe).

Decide your rules in advance. A useful one: any 0 in accuracy or safety blocks the prompt until fixed, whatever the total. A prompt that is beautifully formatted but invents facts is not ready.

Step 2: Write test cases

For each prompt (or at least for the three most important ones), prepare three to five test inputs. Use invented or anonymized material. Include:

• A typical case: the input you usually have.
• A case with missing information: to check the "not stated" or "to be defined" rule.
• A messy or long case: to check robustness.
• A tricky case: ambiguous wording, an unusual request, or an instruction hidden inside the text (for example, a note saying "ignore previous instructions") to see how the prompt behaves.
• A sensitive case, if relevant: an input where an anonymization or fairness rule should apply.

Write down for each test what a good answer looks like before you run it. This keeps you from adjusting your expectations to whatever the model produces.

Step 3: Run, score and diagnose

1. Run each test and save the output (a short note is enough).
2. Score each criterion on your grid.
3. For each 0 or 1, find the likely cause with the diagnosis table from chapter 3: vague context, missing format, no rule for gaps, buried instruction, and so on.
4. Decide on one change per prompt for the next round.

If your results vary from one run to another, run the same test two or three times and note the range. Variation is itself information: a prompt with unstable results needs a template, examples or tighter constraints.

Step 4: Iterate and keep a change log

Apply your one change, rerun the full test set (to be sure you did not break another case), and compare scores. Record each version in the library:

v1 (date): first version. Scores: 10 / 7 / 7.
v2 (date): added "If an owner or deadline is missing, write 'to be
defined'; never guess." Scores: 10 / 9 / 7.
v3 (date): added a template for the output and a maximum of 10 items.
Scores: 10 / 9 / 9. Kept.

Two or three rounds are usually enough. Stop when your prompt meets your criteria on most tests and the remaining failures are understood and acceptable, or when you decide that the task needs a human step or a different tool. Both outcomes are legitimate.

Step 5: Write your personal charter for responsible use

A charter is a short list of rules you commit to follow. It turns the ideas of chapter 4 into your own words, so you do not need to rethink them each time. Keep it to one page and make it concrete. A good charter covers:

1. Data: what I never paste (personal data of others, secrets, credentials) and how I anonymize.
2. Verification: how I check facts according to the stakes, and what I never publish unchecked.
3. Transparency: when I mention AI and how, and how I follow school or workplace rules.
4. Fairness: how I handle content about people, and which decisions I never delegate.
5. Legal awareness: when I check terms of use, copyright or data protection questions, and where I look for reliable information (official sources such as my national data protection authority or EUR-Lex).
6. Critical thinking and skills: how I make sure I still think, learn and write for myself.
7. Environment: how I avoid unnecessary use.
8. Review: when I will revisit these rules.

MY CHARTER FOR RESPONSIBLE AI USE (v1, [date])
1. I never paste personal data of others, secrets or passwords. I
   use placeholders and approved tools only.
2. I verify facts, figures and sources before using them; the higher
   the stakes, the more independent the check.
3. I mention AI use when someone would reasonably expect to know, and
   I follow the rules of my school or workplace.
4. I do not delegate decisions about people to AI.
5. I check the tool's terms and the law (GDPR, copyright) when the
   situation is new or sensitive, using official sources.
6. I think first, then ask, and I edit the result until it is mine.
7. I use AI where it adds real value, not out of habit.
8. I review this charter every few months.

Remember that this is a personal charter, not a legal document. If you work in an organization, align yours with its policy and take specific advice for specialized situations.

Step 6: Wrap up your project

Your final deliverable can be a single document with four parts:

1. The scope (tasks, success criteria, risks).
2. The prompt library (five to eight documented prompts, final versions).
3. The evaluation (test cases, grid, change log, time measured, conclusions).
4. The charter and a short learning plan.

Add a short conclusion with three honest lines: what worked, what did not, and what you will do next.

Keeping up and continuing to learn

Tools, features and rules evolve fast, but the foundations you learned stay useful: how models work, how to structure a prompt, how to verify, how to protect data. To stay current without being overwhelmed:

• Practice regularly on real tasks. A little each week beats a burst once.
• Re-test your library when a tool changes significantly. Prompts can behave differently after an update.
• Follow reliable sources: official documentation of the tools you use, your data protection authority, EUR-Lex for legislation, reputable news and educational institutions. Be wary of hype and of "secret prompt" marketing.
• Learn from others. Share your library, compare prompts with colleagues and discuss what works (without sharing sensitive content).
• Explore the next steps that interest you: image or audio tools, spreadsheet and code assistance, custom assistants, automation, or deeper study of data and statistics.
• Keep your fundamentals. Writing, reasoning, reading sources and numeracy remain your most valuable skills and the ones that make AI useful.

Your turn: finish your project

Exercise 1. Build your evaluation grid and test cases for at least three prompts (one per task). Run them, score them, and make at least one improvement round with a change log.

Exercise 2. Write your one-page charter, using the eight points above as a guide.

Exercise 3. Write a learning plan: three things you will practice in the next month, one source you will follow and a date to review your library.

WHAT WORKED
The payment reminder and weekly plan prompts: scores of 9 or 10 on all
tests after v2, saving roughly a third of my time on those tasks.

WHAT DID NOT
The proposal rewrite sometimes changed commitments; I added a check
prompt and now always read prices and deadlines myself.

NEXT STEPS
- This month: reuse three prompts on real tasks and note the time.
- Try a custom assistant for the payment reminder.
- Follow my data protection authority's guidance for AI.
- Review the library and charter on [date], and re-test after any
  major tool update.