What you will be able to do
- Break a business problem into steps and decide which collaboration mode fits each step: automation, augmentation or agency
- Explain why implementation can be delegated to Claude but judgment cannot
- Write concrete acceptance statements that define done before any work is handed to Claude
- Trace a solution back through the Description Chain, and use Claude to help turn a need into a spec
Key concept
Definition of done (acceptance tests) — A short set of concrete statements, written before anything is built, that let anyone tell whether the finished solution passes or fails. You and Claude both work toward these statements, and every later iteration is measured against them.
1.Start with the problem, not the tool
The AI Fluency Framework breaks working with AI into four connected competencies: Delegation, Description, Discernment and Diligence. When you use Claude to design and build a solution, the first three appear in that order. You decide who does what, you describe it precisely, and you judge whether the result is good enough.
Delegation does not mean asking "should I use AI here?" about the whole project. You begin with the customer problem, break it into steps, and then choose a role for Claude at each step. Anthropic's builder course offers three collaboration modes for that choice:
- Automation: Claude does the work and you check it. - Augmentation: you and Claude work the task together. - Agency: Claude works with latitude inside boundaries you set.
One line runs through all three modes. Implementation is safe to delegate, but judgment is not. Claude can produce the build. Deciding which problem matters, who the users are and what counts as good enough stays with you. As Claude speeds up implementation, more of your value comes from framing the problem and raising the quality bar.
Checkpoint 1 of 4· Match them up
Match each collaboration mode to how the work is shared.
Tap a term, then the definition that fits it.
These are the three modes the builder course uses when you assign a role to AI at each step of a decomposed problem.
“Automation (AI does it, you check), Augmentation (you and AI work it together), or Agency (AI operates with latitude inside boundaries you set)”Source: academy.claude.com
2.Define "done" before anything is delegated
Before Claude builds anything, write down what success looks like. The builder course asks for five to seven acceptance statements. Each one must be concrete enough that a stranger could check the finished product against it and say pass or fail.
The course contrasts two statements. "A patient with a basic smartphone can find the current wait time in under 30 seconds without creating an account" is a good one. It names the user, the device, the time limit and a constraint. "The system is easy to use" is a bad one, because nobody can test it. Two people could disagree about whether it passed, and Claude has no target to aim at.
These statements do two jobs. They tell Claude what to aim for when you hand over the work. They also become the yardstick for every later decision: whether a draft is good enough, whether a change made things better or worse, and whether a pilot should continue. Without them, iteration turns into opinion.
Checkpoint 2 of 4· Check yourself
A support manager is about to have Claude draft replies to routine tickets. Which statement is most useful as a definition of done?
Only this statement is specific enough that someone outside the team could check a draft and say whether it passes. The other three are adjectives and feelings that nobody can test.
“Each must be concrete enough that a stranger could tell you whether the finished product passes or fails.”Source: academy.claude.com
Checkpoint 3 of 4· Exam question
Over three weeks a team refined the instructions for a Claude-drafted internal newsletter until the editor was consistently happy with every draft. Nobody asked the readers. What is the greatest risk?
Correct answer: B — The drafts are now tuned to one editor's taste, and no one has checked that they serve the readers.
- A. Long instructions are an onboarding inconvenience that can be tidied up. They make the arrangement harder to hand over, but they do not make the newsletter less effective for the people reading it.
- B. Correct. Iteration improves whatever it is measured against, and here the only measure was one editor's approval. Three weeks of refinement have optimised for a proxy, so the drafts could satisfy the editor completely while serving readers worse than before.
- C. The refinement effort is a sunk cost that will be repaid if the arrangement works. Time spent iterating is expected; the problem is what the iteration was aimed at.
- D. Concentrating judgment in one person is a genuine weakness in the design, and it follows from the same root cause. It is a consequence of tuning to one taste rather than the primary danger, which is that readers were never consulted.
Sources2
3.From the user's need to the instruction Claude acts on
Once you know what done looks like, you have to describe the work so that Claude can carry it out. The builder course calls this path the Description Chain. It runs from the user's own words, to a requirement, to a technical spec, to the instruction Claude acts on. Prompt writing is only the last link in that chain. You act as the translator at every step, because Claude cannot hear what the user never said. Each step also involves judgment calls that the step before it did not make for you.
Claude helps most in the middle of the chain, in augmentation mode. You share your requirement, ask Claude to propose an approach, and push back on anything that does not fit your constraints. You can also reverse the roles and have Claude interview you. The Claude Code best-practices guide gives a prompt that starts like this:
I want to build [brief description]. Interview me in detail using the AskUserQuestion tool.The prompt goes on to ask Claude to dig into the hard parts you might not have considered. It ends by asking Claude to keep going until everything is covered and then write a complete spec. You review that spec. You do not simply accept it.
One more design rule applies as you write the spec. Anthropic's engineering guidance recommends the simplest solution that works, adding complexity only when it is needed. For many applications, a single well-prompted call is enough, and you do not need a multi-step agent.
Checkpoint 4 of 4· Check yourself
You ask Claude to propose a technical approach for your requirement. In augmentation mode, what is your role once the proposal arrives?
In augmentation mode you and Claude work together. Claude proposes and you push back wherever the proposal conflicts with your constraints.
“share your product requirement, ask it to propose a technical approach, then push back on anything”Source: academy.claude.com
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.If Claude can build the solution, it is just as safe to let Claude make the judgment calls about it.Why is that wrong?
Delegating implementation is safe, but delegating judgment is not. Framing the problem and deciding what counts as good enough stay with a person.
Covered in Start with the problem, not the tool
2.A goal such as "the system is easy to use" is a good enough success criterion to hand to Claude.Why is that wrong?
Acceptance statements must be concrete enough that a stranger could say whether the result passes or fails. Adjectives alone cannot be checked.
Covered in Define "done" before anything is delegated
3.When the solution misses what users need, a better prompt will fix it.Why is that wrong?
The prompt is only one link in the Description Chain. If the user's need was never captured in the requirement, Claude cannot recover it, so the break is usually further upstream.
Covered in From the user's need to the instruction Claude acts on
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“The framework breaks fluency into four interconnected competencies: Delegation, Description, Discernment, and Diligence.”
↩︎ Start with the problem, not the tool - 2.https://academy.claude.com/courses/ai-fluency-for-builders/delegation-the-builder-s-toolkitOfficial docs
“As AI accelerates implementation, your value shifts to framing problems and raising the bar.”
↩︎ Start with the problem, not the tool“A patient with a basic smartphone can find the current wait time in under 30 seconds without creating an account.”
↩︎ Define "done" before anything is delegated“Write acceptance tests before code. They give you and AI a shared definition of done.”
↩︎ Key concept“Explain why delegating implementation is safe and delegating judgment is not”
↩︎ Exam trap 1“Each must be concrete enough that a stranger could tell you whether the finished product passes or fails.”
↩︎ Exam trap 2“Delegation means decomposing the problem first, then deciding what AI handles at each step.”
↩︎ Prediction“Automation (AI does it, you check), Augmentation (you and AI work it together), or Agency (AI operates with latitude inside boundaries you set)”
↩︎ Checkpoint“Each must be concrete enough that a stranger could tell you whether the finished product passes or fails.”
↩︎ Checkpoint - 3.https://academy.claude.com/courses/ai-fluency-for-builders/description-building-great-thingsOfficial docs
“The Description Chain connects user voice to requirement to technical spec to AI instruction. Prompt engineering is only one link.”
↩︎ From the user's need to the instruction Claude acts on“Every step involves judgment calls the previous phase did not make for you.”
↩︎ From the user's need to the instruction Claude acts on“The builder is the translator at every step. AI cannot hear what the user did not say.”
↩︎ Exam trap 3“share your product requirement, ask it to propose a technical approach, then push back on anything”
↩︎ Checkpoint - 4.https://code.claude.com/docs/en/best-practicesOfficial docs
“Keep interviewing until we've covered everything, then write a complete spec to SPEC.md.”
↩︎ From the user's need to the instruction Claude acts on - 5.https://www.anthropic.com/engineering/building-effective-agentsSecondary source
“we recommend finding the simplest solution possible, and only increasing complexity when needed”
↩︎ From the user's need to the instruction Claude acts on