How to develop and assess context management in AI-enabled work

Start with a context budget

Define what the output must help someone decide or do. Then list the minimum information categories needed: task instructions, audience, evidence, constraints, definitions, examples, and output format. For each category, ask whether the system needs the full source, an excerpt, a structured fact, or only a reference for later verification.

NIST's AI RMF Core says intended purpose, context of use, knowledge limits, human oversight, and relevant risks should be understood and documented. NIST also notes that AI RMF 1.0 is being revised. That framework does not prescribe a prompt template, but it supports treating context as an operational boundary rather than a volume target.

The UK government's guidance to civil servants on generative AI warns against entering information that could cause harm if compromised, asks users to consider context a tool may have missed, and requires generated facts to be checked against reliable sources. The guidance is specific to civil-service use. Here it provides a current primary example of why input selection and output verification belong in the same workflow.

Sources: [1] [2]

Build the context manifest

Use six columns:

Add a short exclusion note. Naming what is absent prevents both the user and a reviewer from assuming the model saw an attachment, previous decision, policy exception, or later dataset that was never provided.

• Input: the document, record, instruction, or fact supplied.
• Role: authoritative requirement, supporting evidence, background, example, or constraint.
• Scope: which question, period, audience, or decision the input applies to.
• Freshness: version, date, owner, and whether a newer source might exist.
• Handling: approved tool, sensitivity limit, excerpting rule, and information deliberately excluded.
• Check: how the final output will be traced back, recalculated, or challenged.

Worked artifact: a synthetic supplier-review brief

A fictional operations analyst asks an AI assistant to prepare a quarterly supplier review. The available folder contains a current contract, a service-level report, complaint notes, last quarter's deck, an unsigned renewal draft, and a spreadsheet exported before two late invoices were corrected.

Instead of uploading the folder unchanged, the analyst makes this manifest:

| Input | Role and scope | Handling | Output check | | --- | --- | --- | --- | | Signed contract v4 | Current service commitments | Approved excerpt of relevant clauses | Trace each stated obligation to clause | | Q3 service report | Current operational evidence | Full approved report | Recalculate three headline rates | | Complaint log | Qualitative evidence for Q3 | Remove unrelated personal details | Reconcile counts and label themes as coded judgment | | Q2 review deck | Historical comparison only | Supply two comparison slides | Prevent old targets from being presented as current | | Unsigned renewal draft | Excluded | Do not provide to the tool | Note that future commercial terms are out of scope | | Corrected invoice export | Current cost evidence | Use revised file, preserve revision date | Reconcile total with finance control figure |

The first output says the supplier breached a 99.9 percent target. The trace fails because that target came from the old Q2 deck, while the signed contract uses a different measure. The analyst removes the stale target from the active context, clarifies the metric definition, and regenerates only the affected section. A second check confirms the revised statement against the contract and source calculation.

The useful evidence is not that the analyst wrote a long prompt. It is that they selected inputs by role, excluded an inappropriate draft, detected a version collision, and repaired the output without losing the audit trail. The example is synthetic and does not establish a universal supplier-review method.

Test context with omission and collision probes

Before accepting the output, run two focused tests.

An omission probe asks which missing fact, document, stakeholder, or constraint could reverse the conclusion. If the answer identifies a material gap, obtain the source or narrow the claim. Do not ask the model to invent the missing content.

A collision probe looks for sources that use the same term differently or provide competing dates, thresholds, owners, and versions. Resolve the conflict with an authoritative source or keep the uncertainty explicit. More context does not solve a version conflict unless the task also identifies which source governs.

Finally, sample output claims and trace them to the manifest. Check calculated values independently, review sensitive handling, and confirm that excluded material did not enter through another route. Record accept, revise, hold, or escalate for each material issue.

Develop and assess the skill proportionately

Practice tasks should include a real selection problem: a stale source, a sensitive record, two definitions, or a missing attachment. Ask the learner to submit the manifest and revised output so the review can see selection, not only prose quality.

A bounded assessment can observe task framing, input selection, exclusion discipline, source traceability, version control, and correction behavior. One exercise cannot prove that a person will manage every context well. Report the opportunity presented, the evidence observed, missing evidence, and useful follow-up questions.

Two Prune can create a controlled work simulation and organize task-bounded evidence for human interpretation. It does not monitor all workplace AI use, certify data handling, promise accurate outputs, produce mature workforce benchmarks, or make employment decisions.