Configuration and Knowledge Management·Task 5.3·Bloom: apply·Difficulty 3/5·8 min read·Updated 2026-07-14

Writing Testable Verification Instructions (CCAO-F)

Create effective system-level instructions

SUBy Solomon UdohReviewed by Solomon UdohAI-assisted · human-reviewed
In short
Testable verification instructions make Claude's verification behavior explicit and checkable. "Cite the source document for every factual claim" is a testable rule, and "say I don't know rather than guessing when the documents do not cover something" makes uncertainty explicit rather than hidden. These instructions only function correctly when paired with a knowledge base that actually contains the cited sources, and they reduce the risk of confident-sounding but unsupported claims.

Making verification a checkable behavior

Some of the highest-value standing instructions govern how Claude verifies itself. The CCAO-F exam treats writing these as an apply-level skill, because verification is exactly the kind of behavior that must be specified concretely to work: a vague wish to "be accurate" fails silently, while a testable verification rule actually constrains the output.

Two verification instructions carry most of the weight. "Cite the source document for every factual claim" makes grounding checkable: you can inspect any claim and confirm whether a source is attached. "Say 'I don't know' rather than guessing when the documents do not cover something" makes uncertainty explicit: instead of a confident guess filling a gap, the gap is surfaced. Together they turn verification from an aspiration into behavior you can audit.

Testable verification instructions
Standing instructions that make Claude's verification behavior explicit and checkable. A citation rule such as 'cite the source document for every factual claim' lets grounding be inspected, and an uncertainty rule such as 'say I don't know rather than guessing when the documents do not cover something' surfaces gaps instead of hiding them. These instructions function only when the cited sources exist in the knowledge base, and they reduce confident-sounding but unsupported claims.

The citation rule and the uncertainty rule

The citation rule works because it is testable. "Cite the source document for every factual claim" specifies a behavior you can verify by inspection: for any factual claim in the output, either a source is cited or the rule was broken. That checkability is what makes it effective where a vague "be accurate" is not, and it is a direct application of writing precise rather than vague instructions.

The uncertainty rule addresses the failure the citation rule alone does not. Without an explicit instruction to admit ignorance, a model can produce a confident, well-formed answer even where its sources say nothing, and that confident guess is hard to spot precisely because it looks like every other answer. "Say 'I don't know' when the documents do not cover something" makes the uncertainty visible instead of hidden, so a gap presents as a gap rather than as a fluent but unsupported claim. The two rules together reduce the risk of confident-sounding but unsupported output.

The uncertainty rule does not have to be phrased as a bare "I don't know". For structured or numeric output, the same discipline can be wired into the format itself: an instruction such as "for every figure in a report, state its source, and if a figure is not in the provided data, mark it 'unverified' rather than including it" turns each unsupported value into a visible flag on the page instead of an invisible guess. The mechanism is identical - surface the gap rather than paper over it - but the flag rides inside the deliverable, which is often more useful than a blanket disclaimer when the output is a table or a set of planning numbers.

Verification instructions must be paired with sources

There is a dependency the exam insists on: a citation instruction only functions when the knowledge base actually contains the sources to cite. This is the pairing rule in its most important application. Tell Claude to cite the source document, but leave the knowledge base empty of those documents, and the instruction has nothing valid to cite; it cannot do its job, and it may even invite fabricated citations.

So writing a good verification instruction is only half the work. The other half is ensuring the referenced material is present. A citation rule wired to a well-curated knowledge base produces checkable, grounded output; the same rule with no sources behind it is inert. Verification behavior is a two-part configuration, and both parts have to be in place.

Cite sources
testable: every claim can be checked for grounding
Say 'I don't know'
surfaces gaps instead of confident guessing
Paired
the cited sources must exist in the knowledge base

What the CCAO-F exam trips candidates on

The exam sets two traps. The first is writing a citation requirement without ensuring the underlying source documents are present in the knowledge base. The instruction looks complete, but with no sources to cite it cannot function; the credited answer supplies the documents so the rule has real material. A well-worded citation rule with an empty base is the tell.

The second is omitting an explicit "say I don't know" clause, which allows confident guessing to pass unchecked. A verification setup that requires citations but never tells Claude to admit gaps still lets fluent, unsupported claims through in the places the sources are silent. The credited answer adds the uncertainty rule alongside the citation rule. Both traps reward treating verification as a complete, paired configuration: a testable citation rule, an explicit uncertainty rule, and the sources to back them.

Worked example

A compliance team wants Claude to stop producing confident but unsupported statements. They add the standing instruction 'be rigorous and accurate.' Later they try 'cite the source document for every factual claim' but leave the knowledge base empty. Neither attempt works. Design the verification configuration correctly.

Each attempt fails for a distinct, instructive reason, and the fix combines both rules with real sources.

"Be rigorous and accurate" is vague: it specifies no checkable behavior, so it fails silently and confident-but-unsupported statements keep appearing. The citation attempt is better in kind, a testable rule, but it was left unpaired: with an empty knowledge base there is nothing valid to cite, so the rule cannot function and may even invite invented citations. That is the first trap.

The correct configuration has three parts. First, the testable citation rule: "cite the source document for every factual claim." Second, the explicit uncertainty rule: "say 'I don't know' rather than guessing when the documents do not cover something," which closes the second trap by surfacing gaps instead of letting confident guessing pass. Third, the paired knowledge base: upload the actual compliance source documents so the citation rule has real material to cite. With the two rules made explicit and the sources present, verification becomes checkable and the confident-but-unsupported statements stop, because every claim must be grounded or flagged as unknown.

Common misreadings to avoid

Misconception

Adding 'cite the source document for every factual claim' is enough to make Claude's answers grounded.

What's actually true

The rule only functions when the cited sources actually exist in the knowledge base. Without them there is nothing valid to cite. The citation instruction and the source documents must be configured together.

Misconception

A citation requirement alone prevents confident, unsupported answers.

What's actually true

Where the sources are silent, a confident guess can still pass unless an explicit 'say I don't know' rule surfaces the gap. Pair the citation rule with an uncertainty rule to catch what citation alone misses.

How this shows up on the exam

Domain 5 questions ask how to stop confident, unsupported output. The credited answer combines a testable citation rule, an explicit uncertainty rule, and a knowledge base that actually holds the cited sources. Watch for the trap of a citation rule with an empty base or a setup that never tells Claude to admit gaps.

This apply-level skill draws directly on vague versus precise instructions for the testability and on the pairing rule for the knowledge-base dependency. It is the sharpest instance of the purpose of standing instructions, where a persistent verification discipline holds across every conversation.

Check your understanding

A team wants to stop Claude from producing confident but unsupported statements from its Project knowledge. Which configuration is most effective?

People also ask

How do I make Claude cite its sources?
Write a testable standing instruction such as "cite the source document for every factual claim," and ensure those source documents actually exist in the knowledge base for the rule to cite.
How do I get Claude to say "I don’t know"?
Add an explicit instruction: "say I don’t know rather than guessing when the documents do not cover something." This makes uncertainty explicit instead of letting confident guessing pass unchecked.
Why must verification instructions be paired with a knowledge base?
A citation instruction only functions when the cited sources are present in the knowledge base. Without them there is nothing valid to cite, so the instruction and the documents must be configured together.

Watch and learn

Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.

No videos curated for this concept yet

We are still curating the best official and community videos for this topic.

Official prep for this domain

Anthropic's own free prep module for this part of the syllabus, on the official prep course. Free with an Anthropic Academy sign-in.

References & primary sources

Adaptive study

Master this concept with Archie

Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.

Start studying