No master prompt can replace a solid data foundation

Why the quality of an AI application is often decided before the first prompt is written.

It is tempting to treat every weak AI answer as a prompting problem. When an answer is generic, we expand the role. When details are missing, we lengthen the task. When structure fails, we add formatting rules, examples and review loops. These measures can help, but they eventually reach a hard boundary: a prompt can only work with the foundation the system is allowed to use.

If obsolete policies, conflicting spreadsheets and unlabelled drafts sit together, even an excellent master prompt has limited power. It can request the "valid version", but it cannot reliably identify that version when validity was never recorded. It can ask for current numbers, but it cannot create freshness when timestamps are missing or corrected sources were never ingested.

The prompt directs the task; the data limits the truth

A prompt describes intent, role, procedure and output. The data foundation determines which claims the system can support. This distinction prevents linguistic precision from being mistaken for factual reliability.

A model can produce a convincing answer from a disorganised repository. That is what makes the problem dangerous. Grammar, tone and structure may look professional while the wrong basis was selected. The quality question is not only whether the AI understood the instruction. It is also which information it was permitted to treat as authoritative, and why.

In a small project folder, a brief, meeting notes, a calculation and several presentations may coexist. Experienced team members may intuitively know which file matters. An AI system processes available signals: filenames, content, metadata, similarity and filters. Without status, date, origin and relations, it selects from incomplete evidence. This creates an upstream grounding and quality risk before generation begins.

Data quality is not one score

"Clean data" sounds unambiguous, but quality is purpose-dependent. A complete customer list may be unsuitable for a current campaign if addresses are stale. A correct figure may be misused when it represents an earlier reporting period. A precise meeting record may mislead when it documents an option that was later rejected.

Six questions help define a usable foundation:

1 · Origin: Where did the information come from, and can the source be recovered?

2 · Validity: Is it a draft, reviewed, approved, superseded or archived?

3 · Timeliness: Which period does it cover, and when was it last checked?

4 · Unambiguity: Are terms, units and identifiers consistently defined?

5 · Access: Who or which system may use it, and for what purpose?

6 · Accountability: Who confirms its domain meaning and necessary changes?

These are not administrative ornaments. They translate human project logic into structures that search, retrieval pipelines and agents can enforce.

More material is not automatically more context

When uncertainty rises, teams often upload more folders, emails and presentations. Volume grows, orientation may not. Duplicates can appear repeatedly in retrieval. Similar drafts compete. A semantically close obsolete document may outrank a short current approval.

A practical pattern separates an active workspace from an archive. The active area contains only sufficiently current sources approved for the task. The archive preserves history and evidence but is not injected into every request without filtering. This separation can be implemented through folders or through machine-readable status, versions, validity periods and retrieval filters. Findable must not mean usable in every answer.

The data card as the smallest governance unit

A compact metadata record can already increase reliability. For every important object, a data card records purpose, source, status, owner, protection class, relations and open questions.

id: policy-042
title: Travel expense policy
source: internal policy register
status: approved
valid_from: 2026-04-01
supersedes: policy-031
owner: Finance Operations
ai_use: internal-approved
review_due: 2027-03-01

The card does not prove that the content is flawless. It makes the reason for using it auditable. When a new version appears, its relation to the superseded version can be recorded and retrieval can be updated.

What AI can sensibly take over

AI is useful for scalable preparation. It can classify document types, suggest metadata, identify possible duplicates, flag unusual values and place contradictory statements side by side. These outputs are proposals, not silent decisions about truth.

For critical fields, preserve the original information. An automated change should retain the source value, proposal, method, timestamp and approval. This allows a later investigation to locate an error in the source, transformation, retrieval or generated answer.

The human role shifts away from manually sorting every file. Its higher-value function is to define meaning, limits and approval: which definition applies, which uncertainty is acceptable, and where the system must pause or escalate.

A robust minimum process

Five repeatable steps can prepare data before AI use:

1 · Inventory: register sources, formats and responsibilities.

2 · Qualify: test relevance, validity, timeliness, protection needs and quality.

3 · Curate: select active sources and remove duplicates or superseded versions from default retrieval.

4 · Provide: configure metadata, access, chunking and retrieval for the use case.

5 · Evaluate: use test questions to verify that the right sources are found and claims are supported.

This is not a one-off cleanup. Data changes, systems evolve and new failure modes appear. The foundation therefore needs review dates, change records and controlled retirement of old versions.

The decisive change in perspective

Prompt engineering remains important. A good prompt specifies the goal, limits, procedure and form. Its effect, however, depends on access to an accountable information space. The prompt defines the assignment; the data foundation defines what the assignment may reliably build upon.

Before extending the next master prompt by another page, inspect the material underneath it. Record origin and status. Separate current approvals from historical versions. Define who confirms validity. Then test not only the answer, but also the evidence selected for it.

Sometimes the strongest AI instruction is not another sentence. It is the decision to organise the foundation so that the system no longer has to guess.

Worksheet: Build the data foundation for an AI task

Choose a real use case and three to five related files. Create a data card for each file and decide which sources belong in the active workspace.

1. Define purpose and decision. Describe the specific AI task, the resulting decision or work product, and the most critical consequence of error.

2. Qualify the sources. Record origin, date, status, accountable role, protection class, contradictions and permitted AI use for every file.

3. Curate the active context. Classify each source as active, reference only, archived or blocked. Justify each classification in one sentence.

4. Develop control questions. Create three questions that require the approved sources. Add one unanswerable question for which the system should transparently abstain.

5. Review the result. Score retrieval and generation separately: were the right sources selected, and is every central claim supported by a passage?

All materials to download — the topic overview and the worksheet:

Scope: The NIST AI RMF is voluntary. Article 10 of the EU AI Act applies to high-risk AI systems and is not presented here as a universal obligation for every AI application.

Sources and professional scope