NewAI Content Creation is now live in Early Access
Turning Point Training Instituteby Training Center
Skip to content
0%
Module 5 · Use one MCP teammateLesson 2 of 2
Video lesson

L10 · Keep working when a tool fails

IntermediateVideo lesson
5 min readIntermediate
Downloads & resources
Download this lesson's materials, ready to use.

Skill packs and prompts are provided in English so they work best with the AI tools. Use them as-is.

Keep Working When a Tool Fails

Learning time: about 23.9–35.0 active minutes. Choose ONE presentation: the text (about 5 minutes, estimated) OR the lesson video (3:55, measured). Then 20–30 minutes of practice and checking (estimated). Do not count both presentations. Setup delays and optional extra study are separate; learner timing has not yet been piloted.

Example prompt

Where you are: Module 5, lesson 10. You have a tool card and one bounded research task. Now practise recovering without inventing access or repeating a risky action.

What you will do

Classify a failed tool result, preserve completed work, choose a safe next step and record what remains unverified. You will practise with supplied failure cards; you do not need to break a real connection.

Teaching

A tool failure is not one problem. “Not connected,” “sign-in expired,” “permission denied,” “usage limit reached” and “request timed out” point to different next steps. Read the actual message before asking the assistant to fix everything.

A disconnected service may need configuration. Expired authentication needs the provider's official account flow. A denied action may be outside the account's role or your permission boundary. A usage limit may require waiting or a financial decision. None of these automatically justifies broader access, another account or paid overage.

Timeouts deserve special care. A read-only request that times out may be repeated once when appropriate. An action that changes external state may already have happened even if its response was lost. Before retrying a send, upload, purchase or publication, inspect the destination or operation record. Our course lab avoids those actions, but the distinction matters when you transfer the method to real work.

Separate “the tool failed” from “the project stopped.” You can continue work that does not depend on the missing result. For our kit, source retrieval can fail while the synthetic CSV analysis and local calculator tests remain available. Mark the source decision provisional and keep working on independent tasks.

A saved official-source excerpt can support a bounded claim with its retrieval date. It cannot establish today's product interface, price or account access. Say exactly what the fallback supplies. Never replace a missing live result with a confident answer from memory and describe it as verified.

Use learner/templates/HANDOFF.md to preserve the current state. Include the operation attempted, result, files already completed, unresolved dependency and the next safe action. Do not copy tokens or personal account details into the record. A useful error note is short enough for another person to act on.

Worked example

Faisal's documentation connection returns “authentication expired.” His task is read-only retrieval of GitHub guidance. He records the error, signs in through the official interface when available and makes a small read-only test. Meanwhile, he checks his calculator's zero-attendee case using supplied inputs.

If sign-in remains unavailable, he uses a dated source excerpt for the workflow explanation and labels live retrieval unverified. He does not ask the assistant to search the laptop for credentials. The project continues, but its access record stays honest.

Do it yourself

Use these simulated failure cards, not fabricated accounts of real tool use:

CardObserved resultTask
AAuthentication expiredRetrieve a public documentation page
BTool call timed out; no receipt returnedPost a project update to a remote workspace
CCurrent allowance exhaustedRetrieve a second documentation page
DPermission denied for writeRead a source, then unexpectedly save a change to the provider

For each card, write: what you know, what you do not know, immediate safe step, independent work you can continue and what requires the owner.

Then ask the assistant:

Example prompt

This is a simulated failure-handling exercise for our local Workshop Launch Kit. Card B says a remote update request timed out and no receipt returned. We do not know whether the update happened. Do not call any tool or post anything. Explain the smallest check needed before a retry, the evidence that would permit or prevent retrying, and one project task we can continue while the result is unresolved. Do not assume success or failure.

Inspect whether its answer preserves uncertainty. Save your decisions in the handoff template, clearly marked simulated.

Feedback and retry

Pause here before reading.

A requires official reauthentication or a labelled fallback. B requires checking remote state before any repeat; a missing response is not proof of a missing action. C does not authorize spending or switching billing accounts. D reveals an action beyond the read-only task; return to retrieval without granting write access unless a new purpose is approved.

Retry with a different problem: the tool returns a page, but the quoted claim is absent from it. This is an evidence failure, not a connection failure. Narrow or withdraw the claim and seek the correct source. Reconnecting the tool will not make the source support it.

Check your understanding

Choose an answer and write one sentence explaining why before you read the answers.

  1. Your tool cannot retrieve a current pricing page. A saved screenshot shows a price from last year. Can you call that today's price?
    • A. Yes, prices rarely change.
    • B. Yes, if you add “approximately”.
    • C. No: use it only as dated historical evidence; the current price remains unverified.

Answers

  1. C. No. Use it only as dated historical evidence. The current price remains unverified, and any purchase decision that depends on it must wait for a current check.

Summary

Preserve work, diagnose the actual failure and continue independent tasks. A good handoff names the blocked dependency without turning it into a claim of completed work.

Sources