DEV Community: Daniil Pustovalov

The most useful AI answer can be: not in the document

Daniil Pustovalov — Tue, 13 Oct 2026 14:09:27 +0000

Give an assistant a brand brief and ask a question that the brief does not answer. A fluent response can look like a policy even when it is only a plausible guess.

Here is a small, fictional coffee-brand exercise. The entire source packet says:

Write in plain, welcoming language. Describe the drink and its preparation without health claims. Use close photographs of the drink in natural light. The brand owner must approve product information before publication.

Now ask whether a reseller may print the logo on its own packaging. The document contains no partner-logo policy. “Yes, with clear space around the logo” might sound reasonable, but it is unsupported.

Use a more explicit request: answer only from this packet; quote the supporting sentence for every factual claim; when it does not answer, write “Not in the document” and name the missing input.

The accepted response to the packaging question is: “Not in the document. Ask the owner for the approved partner-logo policy.” The source can support a photography rule. It cannot authorize packaging.

Keep three columns during review: proposed claim, exact source passage and decision. A missing passage becomes a question for the owner rather than a sentence in the final draft. Check quotations yourself; the instruction does not guarantee compliance.

This is an original teaching demonstration, not a client story or a measured model experiment. Its useful result is the evidence map, not a claim that one prompt eliminates hallucinations.

The complete example and downloadable worksheet show how to repeat the check. You can also try a guest AI exercise.

Before debating attribution, reconcile the lead counts

Daniil Pustovalov — Tue, 13 Oct 2026 14:05:24 +0000

Analytics shows 120 successful-form events. The CRM shows 95 leads. Neither total alone tells you where the missing leads went.

Start by agreeing on the unit. An event row, an accepted inquiry, a unique person, a qualified lead and a closed deal are different objects. Choose one cohort and use the original inquiry time consistently.

Here is an original synthetic example, not a client or website result. For October 1–7 in America/Los_Angeles, an accepted-request log has 100 production inquiry IDs. Analytics has 120 rows: 100 first emissions, 15 repeats and five internal tests. The CRM has 95 matching records. A delivery log identifies three queued inquiries and two rejected deliveries.

The reconciliation is:

120 event rows − 15 repeats − 5 tests = 100 comparable inquiries.
95 CRM records + 3 queued + 2 rejected = 100 inquiries accounted for.

That explains the example. It does not prove that queued or rejected inquiries have been delivered. The integration owner still needs to inspect and safely retry them, then verify the destination record.

Preserve the evidence and transformation rule for every subtraction. Do not remove a row simply because it helps two dashboards agree. In real data, consent coverage can also leave the CRM with inquiries that analytics never collected. The bridge can run in either direction.

Keep qualification separate from delivery. A CRM record can be present and unqualified; a person may also make multiple distinct inquiries. State the deduplication rule rather than hiding it in a spreadsheet formula.

The completed reconciliation worksheet includes sources, owners and a repair decision for each group.

Before you publish an AI draft, trace the claim

Daniil Pustovalov — Mon, 12 Oct 2026 18:50:57 +0000

An AI assistant writes: “This coffee cures a headache.” The product brief says only that it is medium-roast coffee from Đắk Lắk. The sentence reads smoothly, but the source does not support its promise.

This is a small example of a review step that belongs in an AI-assisted workflow. A better prompt can improve the draft; it cannot supply missing evidence.

Keep a source beside the draft

Use the approved product sheet, original document or brief. For each factual statement, ask where the evidence comes from. If you cannot point to it, mark the statement as unverified rather than letting fluent wording turn it into a fact.

Separate editing from adding claims

Changing the tone does not require a new promise. “Medium-roast coffee from Đắk Lắk” can be rewritten for clarity without becoming a claim about health or guaranteed results.

Give the reviewer a concrete next action

The next action might be to remove the claim, ask for evidence or flag a missing detail. Saying “make it better” leaves the decision unclear. Saying “remove the unsupported health claim and keep only the source-backed product facts” identifies the problem and the boundary.

Recheck the rewrite

A corrected version can introduce another mistake. Check names, numbers, dates and the call to action again. Record which source you used and why you changed a claim, so a colleague can follow the decision.

Try the exercise

I built an interactive version of this example in my AI Practicum. Try the first task without signing up: compare a coffee draft with the facts, choose a next step and read prepared feedback. This first example does not call a live AI model.

My two AI courses are currently open for free testing: the Practicum has 93 lessons across 13 modules, and AI for marketers has six introductory lessons. A free account saves Practicum progress after lesson one; AI practice responses have usage limits.

Which review decision would you add to this workflow?