Interactive AI Lab | Context Wire

Choose which pieces of information deserve space in a limited prompt. You will see how useful context can beat a larger but unfocused packet.

60-second prompt puzzle

Pack a focused prompt

Pack the highest-value context without crossing the token budget. Tap cards to keep or drop them.

Packet size 9 / 10 tokens

Signal score 28 points

Demo: The essential rules, question, and schema fit with one token to spare.

How the score works

Every card has a small teaching value. The game rewards useful context that fits; crossing the budget applies a penalty. It demonstrates a trade-off, not a universal prompt recipe.

Work through a failed citation scenario one decision at a time. Your choices show how containment, investigation, and a careful return to service fit together.

Three-minute incident story

The citations look real. The evidence is not.

Choose a response at each checkpoint. This tabletop exercise teaches containment, diagnosis, and controlled recovery.

Resolved demoReliability 100

Completed demo: Pause the affected path → rebuild with a freshness gate → return through a monitored canary.

Press “Start scenario” to make your own decisions.

Short demo: the response pattern

Contain: pause the affected path.
Diagnose: test the observed retrieval failure.
Recover: canary the fix with monitoring.

A learning simulation, not an operational guarantee. Real incidents require system-specific authority, evidence, and rollback plans.

Change the number and speed of tool calls, then race two ways of running them. The result shows when independent work can finish faster together.

Architecture race

Serial or parallel: which path finishes first?

Change the number of tool calls and their average duration, then run a local race. The model step is fixed at 600 ms.

Tool calls 3 Each tool 400 ms

Serial chain1,800 ms

S

Parallel fan-out1,000 ms

P

Demo result: Parallel fan-out saves about 800 ms when three independent calls each take 400 ms.

1. Set calls2. Set tool time3. Run race

Educational model only: real latency also includes queues, network variance, retries, and dependent calls that cannot run in parallel.

Move four search results into the order that best answers a question. The score helps you spot direct evidence, supporting detail, and noise.

Retrieval ranking challenge

Put the best evidence first

Query: “Why did my prompt cache stop hitting, and what should I inspect?”

1
Cache key requirements
Cache hits require an identical prefix, including tool definitions and system instructions.
2
Pricing overview
Cached input may use a different price from uncached input, depending on the provider.
3
Debug checklist
Log prefix hashes, inspect ordering, and compare the first changed token when hits disappear.
4
Unrelated release notes
The dashboard added a new color theme and export button this month.

Short demo and answer key

Move “Debug checklist” to position 2, then check the ranking. A useful retriever promotes direct diagnostic evidence and suppresses unrelated text. The exact order can still depend on the task.

Sort short statements by the role they play in a decision. The answer key explains the difference between what was observed, what is only expected, and what someone chose to do.

Three-card challenge

Separate facts, expectations, and choices

Choose the role each statement plays. This is a reasoning warm-up, not a diagnostic score.

Card 1 of 3 0 correct

The test completed 500 requests with a 1.2-second median response time.

Choose a label, then compare your reasoning with the explanation.

Read the complete answer key

Evidence: The test completed 500 requests with a 1.2-second median response time. It reports an observed result and the size of the test.
Assumption: Customers will probably prefer a shorter answer. It is plausible, but it still needs user research or an experiment.
Decision: Release to 10% of users and review errors after one day. It names an action, a limited scope, and a review point.

Adjust speed, control, and budget to compare three operating routes. Presets make it easy to see how a different goal can change the strongest fit.

Interactive decision demo

See how priorities change the route

Move the three priorities or try a mission preset. Scores are illustrative—not product recommendations.

Hosted API63

Fast start, less infrastructure control.

Local runtime52

More control, more operational ownership.

Batch queue58

Budget-friendly when immediate response is unnecessary.

Current illustration: hosted API has the strongest fit at 63 points.

How this demo scores the routes

Each route uses a fixed weighted mix of speed, control, and budget. Real decisions still require measured cost, quality, privacy, and reliability evidence.