LIWLuttrell Intelligence WorksPractical AI training for real work

checklist

Evaluation Checklist

A starter checklist for reviewing AI-generated work before it leaves draft mode.

Use this worksheet before AI-generated work leaves draft mode. It is for a learner, teammate, or reviewer who needs to decide whether an AI draft is ready to use, needs revision, or needs a deeper expert review.

When To Use It

Run this checklist when AI helped create something another person may read, act on, publish, send, or use to make a decision.

Use the short version for low-risk drafts, such as an internal meeting summary or a first-pass outline. Use the full worksheet when the output includes facts, recommendations, instructions, customer-facing language, sensitive data, or any decision that would matter if it were wrong.

Stake-Level Triage

Start by marking the risk level. The goal is not to slow every draft down. The goal is to match the review depth to the cost of being wrong.

Low Stakes

Use a light review when the draft is reversible, internal, and easy to correct.

  • The output is for brainstorming, personal notes, or an early outline.
  • No one will make a business, safety, legal, financial, or customer decision from it.
  • A mistake would be annoying, not damaging.

Review action: read for fit, remove obvious errors, and label anything uncertain.

Medium Stakes

Use a careful review when the draft will shape work, communication, or expectations.

  • The output summarizes source material, explains a process, or recommends next steps.
  • A teammate, stakeholder, learner, or customer may rely on it.
  • A mistake would waste time, create confusion, or weaken trust.

Review action: check important claims against a real source and revise unsupported details.

High Stakes

Use expert review when the draft affects decisions, risk, money, safety, policy, public statements, or production systems.

  • The output includes legal, medical, financial, safety, compliance, security, or HR guidance.
  • It affects a customer commitment, public communication, contract, workflow, or system change.
  • A mistake could create harm, breach trust, expose data, or cause operational damage.

Review action: do not rely on the AI draft alone. Treat it as a starting point and route it to a qualified human owner.

Accuracy And Grounding

Check whether the answer is supported by source material you can inspect.

  • Mark every name, date, number, quote, citation, policy, requirement, file path, command, and recommendation.
  • For each important claim, identify the source that proves it: a document, system of record, meeting notes, policy, code, ticket, or domain expert.
  • If the prompt did not provide the source material, assume the model may have filled gaps.
  • Verify links, citations, document titles, package names, API names, and commands before sharing them.
  • Change confident unsupported claims into draft language, questions, or items to verify.

Ready test: a reviewer can point to where the important facts came from.

Completeness And Fit

Check whether the answer actually solves the task that was asked.

  • Compare the output to the original request, not to what the answer happens to cover.
  • Confirm that required sections, constraints, audience, format, word count, and exclusions are present.
  • Look for silent omissions: skipped edge cases, missing assumptions, absent caveats, or requirements mentioned in the prompt but not answered.
  • Remove extra material that does not help the reader act.
  • Check that the final answer tells the reader what to do next when action is expected.

Ready test: the draft answers the real ask without making the reader infer missing steps.

Hallucination Red Flags

Slow down when the output contains details that are easy to write and hard to verify.

  • It gives precise numbers, dates, titles, laws, versions, or names without showing where they came from.
  • It invents citations, links, quotes, policies, tickets, files, or product features.
  • It describes a system, codebase, decision, meeting, or document the model was not given.
  • It turns uncertainty into confident wording such as “always,” “guaranteed,” “proven,” or “best practice.”
  • It gives step-by-step instructions for tools, APIs, or workflows that you have not tested.
  • It agrees with a false premise instead of challenging it.

Review action: verify the claim, soften it, ask for source material, or remove it.

Tone, Audience, And Ownership

The final work should sound like a responsible person owns it.

  • Match the audience’s knowledge level. Define necessary terms for beginners and remove jargon that does not help.
  • Remove AI boilerplate such as “Certainly,” “As an AI language model,” or a generic closing that does not fit the situation.
  • Check that the tone fits the channel: concise for internal updates, careful for customer messages, plainspoken for training, and formal only when the context requires it.
  • Replace vague praise with useful specifics.
  • Make sure any recommendation is owned by a person or team, not presented as something “the AI decided.”

Ready test: the draft reads like something the sender would stand behind.

Data And Confidentiality

Review both what went into the prompt and what came out of the model.

  • Confirm the output does not expose secrets, credentials, private keys, tokens, internal-only URLs, or security details.
  • Remove or generalize personal, customer, employee, financial, contract, and proprietary business details unless their use is approved for this context.
  • Check whether source material was copied into the output when a summary would be safer.
  • Do not include private messages, email contents, client records, or internal records in a public or broadly shared draft.
  • If the output must keep sensitive detail, confirm the audience, channel, and storage location are allowed for that data.

Ready test: the draft can be shared with the intended audience without exposing data that should have stayed private.

Sign-Off Block

Use this block before sending, publishing, or acting on a reviewed AI draft.

  • Reviewer:
  • Date:
  • Draft owner:
  • Intended audience:
  • Stake level: Low / Medium / High
  • Decision: Ready / Revise / Needs expert review
  • Sources checked:
  • Claims still uncertain:
  • Data or confidentiality concerns:
  • Notes for the next reviewer:

Compact Copyable Checklist

Copy this version into a document, ticket, or message when you need a quick review pass.

- [ ] I marked the stake level: low, medium, or high.
- [ ] I identified who will read or act on the draft.
- [ ] I marked names, dates, numbers, quotes, citations, policies, commands, and recommendations.
- [ ] I checked important claims against a real source or labeled them as unverified.
- [ ] I removed invented or unverifiable links, citations, quotes, files, and product details.
- [ ] I compared the answer to the original request and confirmed it covers the required task.
- [ ] I checked that constraints, exclusions, audience, and format requirements were followed.
- [ ] I removed extra material that distracts from the reader's next action.
- [ ] I looked for hallucination red flags: confident specifics, missing sources, or untested steps.
- [ ] I adjusted the tone for the audience and removed generic AI boilerplate.
- [ ] I confirmed a human owner is responsible for the final message, decision, or action.
- [ ] I removed secrets, credentials, private records, and unnecessary sensitive details.
- [ ] I confirmed the sharing channel is appropriate for any remaining sensitive information.
- [ ] I routed high-stakes content to a qualified reviewer instead of treating the AI draft as final.
- [ ] I completed the sign-off block or recorded why the draft is staying in revision.