Skip to content

Decision Sheet

A scannable reference for every AI session. Print it, pin it next to your monitor, or keep it open in a tab. Each section links to the full page if you need detail.

Essential

This is a checklist, not a comprehensive guide. If any item raises a question you cannot answer, follow the link to the relevant Essentials page.

Before you start

Data governance --- full guide

  • What am I uploading? Classify the data before you share it with any AI tool.
    • Published / public domain material? Proceed.
    • My own unpublished work? Safe on paid tiers. Check institutional policy.
    • Personal data about identifiable people? Anonymise first. Check institutional policy.
    • Data under embargo, NDA, or restriction? Do not upload.
    • Student data? Check GDPR / FERPA compliance.
    • Unsure? Do not upload until you have checked.
  • Which tier am I on? Free tiers may use your data for training. Use paid tiers for anything beyond casual experimentation.
  • Have I checked my institution's AI policy? Know the rules before you start, not after.

During your session

Prompting --- full guide

  • Be specific. State: audience, format, length, register, scope, purpose.
  • Provide context. Discipline, topic, expertise level, constraints. Do not assume the model knows your field.
  • Give a role (when helpful). "You are an expert in..." sets expectations about depth and register.
  • Structure complex prompts. Use sections: Task, Context, Requirements, Output format.
  • Provide an example if you want a specific format or style.
  • Iterate, do not restart. Refine with specific follow-up instructions rather than starting over.
  • Ask for step-by-step reasoning on complex analytical tasks.
  • Ask for alternatives. "Give me three approaches..." surfaces more useful material than a single answer.

During use --- good habits

  • Note the model and date for anything you might use later.
  • Save prompts and outputs for work you may publish or teach from.
  • Ask the model to flag uncertainty. "Tell me how confident you are in each claim."

After your session

Verification --- full guide

  • Check every citation. Search library catalogues or databases. If you cannot find it, it may not exist.
  • Cross-reference factual claims against at least one authoritative source.
  • For translations, compare with published translations or check against the original text.
  • For data transformations, spot-check across the full range --- not just the first few items.
  • For code, run it and test the output. Check edge cases.
  • Match effort to stakes. Brainstorming list? Glance. Published claim? Full check. See the verification ladder.

Citation and provenance --- full guide

  • Record: date, model, interface, prompt summary, what you used, how you verified it.
  • For high-stakes work, consider a second model from a different provider. See Multi-Model Strategy.

Disclosure --- full guide

  • Will you publish or submit this work? Check journal / publisher AI disclosure requirements.
  • Draft a disclosure statement while the details are fresh:
    • What model did you use?
    • For what specific purposes?
    • How did you verify the outputs?
    • Who takes responsibility for the content?
  • Apply the "so what?" test. Would knowing about your AI use change how someone evaluates the work? If yes, disclose.

Quick reference: what LLMs get wrong

Keep these failure modes in mind during every session:

Failure mode What to watch for
Fabricated references Plausible but fictional citations. Always verify.
Confident factual errors Wrong dates, misattributions, confused details --- stated with certainty.
Anachronistic reasoning Modern categories applied to historical contexts.
Smoothed complexity Unresolved debates presented as consensus. Ask: "Is there disagreement?"
Translation errors Higher error rates for ancient, low-resource, or specialist languages.

Quick reference: cost awareness

  • Right model for the task. Do not use the most expensive model for simple formatting.
  • Clear prompts save money. A precise prompt resolves in one round; a vague one takes several.
  • Batch similar tasks. One session processing 50 items is cheaper than 50 separate sessions.
  • Monitor usage on subscription plans. Heavy sessions reduce your remaining allocation.

See Cost & Plans for full details.

Don't Panic

This looks like a lot. In practice, most of it becomes second nature after a few sessions. The three things that matter most: (1) check your data before uploading, (2) verify outputs before using them, (3) disclose your AI use. Everything else is refinement.