Limit an AI agent’s permissions before giving it tools
Separate reading from writing, restrict accounts and destinations and require a clear authorization path for consequential actions.
docs.anthropic.com
Find a guide for the task at hand 6
Search for a problem or tool, or browse guides and comparisons by section.
274 topics in Arabic and English
Answer from documents with RAG without treating retrieval as proof
Retrieve relevant permitted passages, attach source references and allow the answer to say when the documents do not support a claim.
OpenAI — Retrieval
Use embeddings for similarity search without confusing similarity with truth
Compare semantic retrieval with keyword search, preserve document permissions and evaluate whether the relevant evidence is returned.
OpenAI — Embeddings
Split documents for AI without cutting the answer away from its conditions
Keep headings, clauses and table context together, then test whether a retrieved chunk can answer the question on its own.
OpenAI — Retrieval and chunking
Make AI citations point to evidence, not just a source name
Preserve source IDs and page references, attach citations to individual claims and verify that the cited passage actually supports them.
OpenAI — Retrieval
Use prompt caching for repeated context and verify the cache hit
Keep reusable material stable, follow the selected model’s caching controls and check usage instead of assuming every repeat is discounted.
OpenAI — Prompt caching
Use Batch API for work that can wait, not an interactive response
Prepare independent tasks with stable identifiers, track job status and recover partial failures without resubmitting successful items.
OpenAI — Batch API
Stream AI responses without treating partial text as a final answer
Handle documented stream events, show an unfinished state and wait for completion before parsing or acting on the result.
OpenAI — Streaming responses
Find the AI latency bottleneck before changing models
Measure queueing, retrieval, first output and completion separately, then reduce the stage that is actually delaying the task.
OpenAI — Latency optimization
Evaluate an AI change with real tasks and visible failure criteria
Keep a representative test set, define what counts as success and compare a new prompt or model without changing the test at the same time.
OpenAI — Evals
Test an AI agent’s failures before connecting real accounts
Use harmless fixtures for tool timeouts, conflicting evidence and denied actions, then verify that the application stops without inventing success.
OpenAI — Safety best practices
Protect an AI API key and bound the requests that can spend it
Keep the provider key server-side, authorize users and limit request size, model access and usage before exposing an AI feature.
platform.openai.com
Handle AI rate limits with a queue and bounded retries
Read the actual error and limit dimension, pace concurrent work and avoid retry storms that make a temporary limit worse.
OpenAI — Rate limits
Estimate AI API cost from usage, not the length of the visible answer
Separate input, cached input, output and tool charges, use current rates and include retries in a realistic workload estimate.
OpenAI — API pricing
Route AI tasks by tested requirements instead of model prestige
Define easy and hard task classes, evaluate candidates on the same examples and use a controlled escalation path.
OpenAI — Model selection
A large context window does not guarantee a complete answer
Understand input and output budgets, keep the needed evidence findable and test whether key conditions survive a long conversation.
docs.anthropic.com
Ask AI about a PDF with page-based evidence, not a vague upload
Check whether the file is text or a scan, ask a defined question and verify tables, conditions and references against the original pages.
OpenAI — File inputs
Use AI vision for screenshots and charts, then verify the detail
Provide a readable relevant image, separate what is visible from interpretation and check small text, units and labels yourself.
OpenAI — Images and vision