Ask AI to identify what you could provide to improve the work. A useful answer names a specific contribution and explains which error, uncertainty, or quality gap it would address. That gives you a practical way to help without reviewing every sentence or guessing what context the model needs.
The earlier posts establish the work product, its specification, its verification plan, and a bounded improvement loop. This post adds a way to obtain missing human input. It also works independently when you have a single draft or analysis to improve.
Ask for a contribution with a purpose
The specification interview clarifies what you want. This prompt examines what the AI needs to achieve it. The answer might be an authoritative record, an annotated example, a site observation, or a decision from the responsible person.
Anthropic’s guidance on context engineering recommends selecting a compact set of relevant information and representative examples for the task [1]. Applied here, that suggests asking for a small, justified contribution. Uploading an entire project archive can leave the AI with the same unresolved question and more material to search.
Require the AI to inspect what it already has. If the answer is available through approved sources or tools, it should retrieve it. Human effort is most useful where access, experience, or decision authority is missing.
The prompt
Use this with the work product or task description, the available sources, and any specification or verification findings.
Review the task, available material, and any verification findings. What is the
smallest contribution I could provide that would most improve the accuracy or
quality of this work product?
First use what you already have and can retrieve within your permissions.
Prioritize requests by their likely effect on the result and the effort they
require from me. Tie each request to a specific gap, explain why it matters,
and say how you would check whether it helped. Distinguish essential input
from optional improvements. Do not invent gaps or numerical confidence gains.
Ask for the highest-value contribution first, in a form I can readily provide.
State what you can continue without it and what must wait. If no material
human input is needed, say so.
After I respond, check the input against the approved sources and requirements,
flag conflicts, and use it only within the agreed scope. Verify the revised
work and report what improved or remains unresolved.Make the request easy to answer
Consider an AI preparing a schedule narrative. Its verification report finds two different completion dates in the supplied material. Another rewrite cannot establish which date governs.
A useful request would identify both dates and their sources, then ask for the approved baseline revision or the planner’s recorded decision. It would explain that the answer determines which milestone the narrative should use. Verification would then check every affected date and the explanation of any variance.
The contribution could also concern usefulness. If a draft satisfies the factual checks but misses the audience’s expected detail, one annotated paragraph from an approved example may clarify what to retain, remove, or explain. As in 004, an example establishes presentation expectations; current project evidence supports factual claims.
Human input still needs interpretation. A recollection may need a source, and two reviewers may disagree. The AI should preserve those distinctions. Research on human-AI interaction supports making correction easy and seeking clarification when the user’s goals are uncertain [2]. The prompt applies those principles to a concrete work product.
Insert the request into the loop
Run the input review before the first iteration, while the sources and controls are being assembled. Then let the bounded loop in 005 proceed through production, verification, selection, and lessons.
After each iteration, the lesson can identify whether the next useful step depends on human input. Optional requests can accumulate until the run ends. A material UNRESOLVED finding still means HALT. An agreed plateau ends the run and creates an opportunity to ask what could enable further progress.
Add this companion instruction to the loop prompt before starting:
Before the run, apply the human-input prompt to identify essential missing
contributions. Resolve them before fixing the controls and starting iterations.
During the run, use each verification report and lesson to identify whether
human input would materially improve the next attempt. Record optional requests
and continue within the approved limits. Do not interrupt routine iterations.
When a material finding is UNRESOLVED, HALT. At a HALT or the agreed plateau,
return the best retained version and the highest-value human-input request,
explaining the gap, the contribution needed, and the check it would enable.
Record any response and its source. If using it changes approved sources,
requirements, verification rules, or other fixed controls, obtain approval for
new versions and start a new bounded run. Otherwise, recheck the affected
conditions and resume only with human authorization and the remaining budget.
Verify the next candidate and record whether the contribution helped.This preserves the existing roles: specification defines the target, verification evaluates the evidence, and the loop selects candidates. The new instruction makes requests at the human boundary more useful.
Check whether the contribution helped
Compare the revised work against the same acceptance conditions wherever possible. Report whether the missing evidence is now available, the inconsistency is resolved, or the agreed quality measure improved. A more confident answer alone is insufficient.
NIST’s AI Risk Management Framework calls for expert and user input to inform evaluation and for resulting improvements or declines to be documented [3]. If the human response changes the criteria themselves, establish a new baseline before claiming improvement across runs.
Optional extensions
For complex work, specify a human time budget, the person authorized to answer, or an approved format for observations. Add only what the project needs. A short request that can be answered precisely is easier to act on than a broad questionnaire.
Questions
Can AI reliably identify everything it is missing?
No. Treat its requests as proposals grounded in visible gaps. Independent verification and domain review remain necessary because an AI may overlook a problem or ask for irrelevant information.
Notes
Join the EPM Network to access insights, influence our research, and connect with a community shaping the industry’s future.
Support us by sharing this article with your friends and colleagues, or over social media.
References
1. Anthropic. (2025, September 29). Effective context engineering for AI agents.
2. Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi, B., Collisson, P., Suh, J., Iqbal, S., Bennett, P. N., Inkpen, K., Teevan, J., Kikin-Gil, R., & Horvitz, E. (2019). Guidelines for human-AI interaction. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, Article 3, 1–13.
3. National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework Core.