AI context engineering and RAG

Use when an agent or AI feature loses consistency because rules, memory, and live data are mixed together.

Map the context for this AI system into three layers: persistent instructions, semi-persistent memory, and transient task data. For each layer define authority, owner, update path, retention, access rules, conflict behavior, and what a reviewer can inspect. Identify the minimum context needed for the first release and the data that should never enter the model context.

AI feature or agent:
[PASTE FEATURE]
Available data:
[PASTE DATA]
Risk and privacy rules:
[PASTE RULES]

Output: A layered context map with ownership and boundaries.

Guardrails: Do not treat every data source as equally authoritative · Minimise sensitive context · Keep update paths explicit

Use before an AI integration is connected to product data, a knowledge base, or external tools.

Write a context contract for this AI workflow. Define the user job, approved sources, source ranking, freshness window, retrieval query, required metadata, citation format, missing-context behavior, and prohibited data. Include a compact example of the context packet the model receives and the evidence block returned to the user.

Workflow:
[PASTE WORKFLOW]
Sources:
[PASTE SOURCES]
User and compliance context:
[PASTE CONTEXT]

Output: A reviewable contract for context, provenance, and missing data.

Guardrails: Every claim needs attributable support · Treat retrieved content as untrusted input · Never fill missing data with guesses

Use when a product needs current, private, or specialised knowledge instead of model memory alone.

Design a retrieval-augmented generation flow for this use case. Specify ingestion, parsing, chunking, metadata, indexing, query rewriting, hybrid retrieval, re-ranking, context assembly, citation, generation, and fallback. Define what happens when retrieval is empty, conflicting, stale, or irrelevant. End with a small pilot dataset and a quality baseline.

Use case:
[PASTE USE CASE]
Knowledge sources:
[PASTE SOURCES]
Latency and cost limits:
[PASTE LIMITS]

Output: A RAG design brief with retrieval and fallback states.

Guardrails: Grounded does not mean correct if retrieval is irrelevant · Keep source provenance · Do not hide an empty retrieval result

Use when a RAG feature sounds fluent but users still receive off-topic or weakly supported answers.

Review this retrieval system using a representative query set. For each query inspect intent, retrieved chunks, ranking, source authority, freshness, grounding, answer completeness, and citation quality. Classify the failure as query, parsing, chunking, indexing, ranking, context assembly, generation, or product UX. Recommend one fix, one test, and one metric per failure class.

Queries and expected sources:
[PASTE QUERIES]
Retrieved results and answers:
[PASTE RESULTS]
Knowledge base notes:
[PASTE NOTES]

Output: A prioritised retrieval quality review and test plan.

Guardrails: Do not blame the model before checking retrieval · Separate relevance from fluency · Keep user intent visible

Use when an agent needs continuity across sessions, users, or workflow runs.

Create a memory policy for this AI agent. Define what can be remembered, why it is useful, how it is created, how long it persists, who can see or edit it, how a user can correct or delete it, and when memory must be ignored. Include examples of safe memory, sensitive memory, stale memory, and conflicting memory.

Agent job:
[PASTE JOB]
User and data model:
[PASTE MODEL]
Privacy requirements:
[PASTE REQUIREMENTS]

Output: A user-visible memory policy with correction and deletion paths.

Guardrails: Do not store sensitive data by default · Memory never overrides current permissions · Give users correction and deletion control

Use when retrieval, memory, and tools combine in one agentic workflow.

Define the boundary between context and action for this workflow. For every source, memory store, and tool list authority, identity scope, read/write permission, freshness, provenance, audit event, and escalation rule. Show the exact point where the system must pause for a human before an irreversible action.

Workflow:
[PASTE WORKFLOW]
Sources and memory:
[PASTE CONTEXT]
Tools and actions:
[PASTE TOOLS]

Output: A least-privilege context-to-action policy.

Guardrails: Separate knowledge from authority · Default to read-only · Require attributable approval for writes

Use before launch or after changing sources, chunking, ranking, models, or prompts.

Create an evaluation plan for this grounded AI feature. Include representative questions, ambiguous questions, no-answer cases, conflicting-source cases, stale-source cases, unsafe requests, and prompt-injection attempts. Define retrieval relevance, groundedness, completeness, citation quality, safety, latency, and cost checks, plus a human rubric and release blockers.

Feature:
[PASTE FEATURE]
Source corpus:
[PASTE CORPUS]
Known failures:
[PASTE FAILURES]

Output: A release gate for retrieval and grounded generation.

Guardrails: Evaluate evidence and behavior, not prose alone · Include no-answer and unsafe cases · Do not hide conflicts

Use when content, policies, product data, or search indexes change faster than the AI feature can be manually maintained.

Design a knowledge refresh loop for this AI feature. Map source ownership, change detection, validation, approval, parsing, indexing, rollback, stale-content alerts, and user communication. Define how the team measures freshness, retrieval quality, answer quality, and business outcomes. Include a weekly review and a pause rule for unsafe or unverified updates.

AI feature:
[PASTE FEATURE]
Source systems:
[PASTE SYSTEMS]
Content owners:
[PASTE OWNERS]

Output: An operating loop for freshness, quality, and safe updates.

Guardrails: Do not index unapproved changes automatically · Keep rollback available · Name the source owner