How to Diagnose AI That Misses Project Decisions

Key takeaways

• Context engineering means designing and managing the information an AI receives, in the right form and at the right time.

• Compare the original decision, retrieved candidates, summary, and actual model input.

• A longer prompt is not a universal repair.

When an assistant contradicts a settled decision, rewriting the prompt is not always the first move. Trace the decision through retrieval, summarization, and the final input. Finding where it disappeared narrows the repair.

For the underlying concept, see What is context engineering? in the AI glossary.

Capture one reproducible failure

Suppose a fictional development team agrees to preserve its payment architecture, then receives a redesign proposal in a later conversation. Keep the question, expected behavior, decision record, and input together. Describe the missed constraint and timing rather than simply saying the assistant forgot.

Check whether the decision reached the input

Context engineering considers supplied material, state, and tool information alongside instructions. If the decision is absent, inspect retention and delivery. If it is present but ignored, inspect precedence and application. Changing the model without this distinction makes the outcome harder to interpret.

Inspect summaries and retrieval results

A summary may retain the conclusion but drop an exception. Retrieval may find a similarly worded decision from another project. Compare constraints before and after summarization, project scope in retrieved candidates, and anything omitted from final input. Keep only necessary diagnostic records in an approved environment.

Change one factor and rerun the case

Test a targeted change, such as recording status and scope or revising input placement. Include paraphrases, revised decisions, and withdrawn decisions. Keep failed conditions as regression cases rather than collecting only successful answers.

A worked business example

Consider a engineering team agreed to preserve the payment architecture, but the assistant proposes changing it. A decision in chat history is not proof that it reached the current request.

Compare the original decision, retrieved candidates, summary, and actual model input. If retrieval succeeded but summarization removed the prohibition, repair input construction before replacing retrieval and repeat the same question.

Failure conditions to check

A longer prompt is not a universal repair. Without locating the missing constraint, added input can leave the original failure unresolved.

A checklist for your workflow

Frequently asked questions

Will repeating an instruction fix it?

It may help, but does not demonstrate a fix for retrieval, summarization, or authorization faults.

Related AI concepts

• What is context engineering?

• What is context in AI?

• What is retrieval-augmented generation (RAG)?

Sources and editorial information

Anthropic: Effective context engineering

TENCONIC editorial team · Structure reviewed October 28, 2026. Official sources explain concepts and implementation examples; features vary by product. Scenarios and checks are editorial recommendations, not customer results or product promises.