# The Capability/Limitation Cheat Sheet

One page: what AI is reliably good at, what it reliably fails at, and what to do about each.

## The rule that predicts most of this page

Closed-book vs open-book. A model answering from memory is closed-book: fluent, shaky on specifics. Paste your source material into the prompt and it becomes open-book: accuracy jumps. **When in doubt, supply the source.**

## Reliably good at

| Task | Why it works | Your move |
|---|---|---|
| First drafts: emails, memos, lesson plans, donor letters, cover letters | Fluent text is the model's core skill | Edit for voice; fact-check anything specific |
| Summarizing documents you paste in | Open-book | Skim summary against the original |
| Reformatting: notes to outline, table to prose, long to short | No new facts required | Quick read-through |
| Brainstorming options, names, structures, counterarguments | Volume and variety, not truth | Pick and refine; discard freely |
| Explaining a concept at a chosen level | Strong on well-documented topics | Verify before you teach or advise from it |
| Critiquing your draft, role-playing a skeptical reader | It reacts to text you supply | Treat as one opinion, not a verdict |
| Rewriting for tone, length, or audience | Transformation of supplied text | Confirm nothing important got cut |

## Reliably fails at

| Task | Failure | Your move |
|---|---|---|
| Citations: cases, papers, statutes, studies | Fabricates plausible references | Verify every citation at the source. No exceptions |
| Current data: rates, deadlines, prices, rules | Training cutoff; knowledge is stale | Check the official source for the current year |
| Precise math and counting | Miscounts, drops digits | Use a calculator or spreadsheet; or have it write the formula and check that |
| Knowing what it doesn't know | States guesses as facts, same tone | Never judge reliability from confidence |
| Your private context it was never given | Invents details to fill gaps | Put the context in the prompt (Stage 1 inventory helps) |
| Long documents in one pass | Drops instructions, drifts mid-document | Work in sections; re-check the middle |

## One failure to expect in your field

- **Attorney:** fabricated case law that cites real reporters
- **Accountant:** last year's thresholds presented as current
- **Teacher:** invented quotations and misattributed sources in materials
- **Military spouse freelancer:** confident wrong claims about a client's product
- **Nonprofit:** precise-sounding local statistics with no source
- **Veteran in transition:** outdated benefit rates and program rules

## Make it yours

For your top three tasks from the Stage 1 leverage map:

1. Task: [TASK] — Open-book or closed-book? [O/C] — What I must verify every time: [LIST]
2. Task: [TASK] — Open-book or closed-book? [O/C] — What I must verify every time: [LIST]
3. Task: [TASK] — Open-book or closed-book? [O/C] — What I must verify every time: [LIST]

Part of the ADAPT framework — vetslearn.com/framework