AI Fluency 4D: Delegating, Describing, Discerning and Taking Responsibility When Working With AI

The earlier posts covered what to know about LLMs, data, evaluation, security and regulation. This one is about how to work with AI. The reference is Anthropic's AI Fluency 4D framework: how its four competencies show up as behavior in an actual conversation, and what to change when the model you work with has a limited number of questions and a short input. It does not cover the full 4D course or how to use any particular tool. I got every practice question right, so there are no misses to note this time either.

1. The four competencies

Anthropic Academy splits working with AI into four competencies.

The same course splits ways of working with AI into three: Automation, where AI executes specific tasks you instruct; Augmentation, where you and AI think and execute together; and Agency, where you shape AI's knowledge and behavior rather than specific actions and let it work independently. The four competencies apply to all three.

2. Delegation: keep the criteria and the final call

Delegation is neither handing everything to AI nor avoiding it. It is splitting the work into pieces and deciding who does each one. Take picking which machines to service from 2,000 lines of equipment logs.

3. Description: write so a newcomer could follow it

Anthropic's prompting guide gives one test: show your prompt to a colleague with minimal context and ask them to follow it. If they would be confused, the model will be too.

Adding the pieces one by one to "analyze this data" gives this.

<logs>
(sensor logs per machine)
</logs>
I'm an equipment engineer deciding which machines to service at today's meeting.
We have two technicians, so we can only service three machines a day; I need the three most urgent.
Analyze the logs above.
Give the result as a table: machine ID | anomaly | source log line | priority (1-3)
Don't write anything that isn't in the logs. If you can't find evidence, write "needs checking".

It has the goal and reader, the reason, the material, the output format and an evidence rule. The guide explains that giving the reason lets the model generalize toward the goal. Three more things to keep in mind:

• Put long material at the top and the question at the end. According to the guide, putting the query last on long inputs improved response quality by up to 30 percent in tests.
• To pin down a format, include three to five relevant, varied examples wrapped in tags.
• Phrases like "be careful, this has to be right" carry no information. Say specifically what to keep.

4. Discernment: check instead of trusting

These are the methods Anthropic's guide to reducing hallucinations recommends.

The guide notes that these methods greatly reduce hallucinations but do not eliminate them. The last one only tells you the answers differ, not which one is right, so asking the same AI again is not verification. In the end you check against the source.

Recompute numbers yourself. If AI says the mean of 10, 12, 15, 9 and 14 is 12.5: 10 + 12 + 15 + 9 + 14 = 60, and 60 ÷ 5 = 12.0, so it was wrong. A conversation that records you catching this is what evidence of verification looks like.

5. Diligence: filter what goes in, own what comes out

• Remove or mask customer names, personal data and unreleased figures before putting material into AI.
• Use only tools your company has approved.
• Don't give unethical instructions or instructions meant to get around rules.
• Responsibility for what you hand in is yours, not the AI's. That is why section 4 matters.

The details are in the security post's section on sensitive information and the ethics and regulation post's section on the user's responsibility.

6. Working with a heavily constrained model

Sometimes you have to work with a model that allows a fixed number of questions, cuts off long input and doesn't remember the earlier conversation. Then the person has to hold the context. Here is the flow on a practice task I made.

There are two files: a spec and chip measurements.

[Shipping criteria]
1. Leakage current: 5.0 µA or less
2. Operating voltage: at least 1.10 V and at most 1.30 V
3. 85°C high-temperature test: must pass (Y)
Exception: for Lot L2, the lower voltage limit is 1.05 V.

id,lot,leak_uA,volt_V,temp85
C01,L1,3.2,1.20,Y
C02,L1,5.4,1.22,Y
C03,L2,4.1,1.07,Y
C04,L1,2.9,1.08,Y
C05,L2,4.8,1.31,Y
C06,L1,1.7,1.15,N
C07,L2,5.0,1.25,Y
C08,L1,3.3,1.19,Y

There are four questions: the IDs of failing chips, the reason each one fails, the failure rate, and which chips would also fail without the L2 exception.

Read all four questions first. Questions 1 and 2 come from the same verdict table, question 3 is arithmetic, and question 4 hangs on the exception. Mark the "or less" boundary too.

Restate the criteria in your own words. Attaching several files can mean only some get read, and missing the one exception line breaks both question 1 and question 4. Spell out the boundary as "equal passes".

[Rules] 1. Use only what is in the conditions below and the attachment. Otherwise "needs checking".
2. Don't calculate. 3. One table, no explanation, only the rows needed to answer.
[Conditions] Leakage passes at 5.0 or less (5.0 passes). Voltage passes from 1.10 to 1.30 (both ends included).
For lot L2 the lower voltage limit is 1.05. temp85 must be Y to pass.
[Task] Give a verdict for each chip in chips.csv.
[Format] id | lot | leakage | voltage | high-temp | final (pass/fail) | reason
Attachment: chips.csv

Get a verdict table with a column per criterion. A final verdict alone gives you no way to check why. Keep answers short; long ones get cut off, and asking the model to continue can lose the context.

Check only the rows on a boundary or exception against the source. C07, with leakage exactly 5.0, passes; C03, an L2 chip at 1.07 V, passes; C05, above the upper limit, fails; C06, failing only the heat test, fails. If a row is wrong, don't ask again; fix it yourself and write the answer.

Do the arithmetic and the exception question yourself. The failing chips are C02, C04, C05 and C06, four of eight, so the failure rate is 4 ÷ 8 = 50%. Without the exception the L2 lower limit becomes 1.10, so the only new failure is C03. If you want the AI to confirm, delete just the exception sentence from the condition card and ask again. Since the model doesn't remember the earlier conversation, put the conditions back in from scratch every time.

This flow needs one or two AI questions. The rest — reading the structure of the questions, restating the criteria, checking boundary rows and doing the arithmetic — is done by the person.

Practice: pick the 4D behavior

1. You need to find suspicious lines in 2,000 lines of equipment logs and set service priorities. What is the best split?

1. You want to ask about service priorities from a dataset. What is the best first question?

1. You want to ask a question about a 50-page process report. Where should things go?

1. AI says the mean of 10, 12, 15, 9 and 14 is 12.5. What is the right response?

1. How do you cut down unsupported claims in a summary of a long report?

1. You have to do a four-question analysis with a model that forgets earlier conditions and attachments and cuts off long answers. What is the best way to proceed?

1. If the conversation log is what gets evaluated, which behavior looks best?

1. The job requires putting material with a customer's name and yield figures into AI. What is right?

Recap

• 4D is Delegation, Description, Discernment and Diligence. The four apply to Automation, Augmentation and Agency alike.
• Delegation means breaking the work up. The criteria, the final call and checking the arithmetic stay with the person.
• A good question is one a newcomer could follow. It carries the goal, context, material, format and rules, with long material on top and the question at the end.
• Verification means asking for quotes, retracting unsupported claims, restricting to the material and recomputing numbers. Asking the same AI again is not verification.
• Mask sensitive details before putting material in, and the result is your responsibility.
• With a heavily constrained model, read the structure of the questions first, restate the criteria yourself, get one verdict table, check the boundary rows and do the arithmetic yourself.