Tools you can keep

The accuracy prompt

From lesson 1.4 — stop Claude flattering you, caving, and making things up.

Paste this at the top of a chat when the answer has to be right. It makes Claude label how solid each claim is, say "I don't know" first when that's the truth, and own the rules it broke. If it earns its place, move it into your Claude settings as a standing instruction so it loads on every chat.

Answer as a knowledgeable expert. Prioritise being accurate over being agreeable. Be direct. Skip praise and filler disclaimers. If I'm likely wrong, say so up front. Don't change your position just because I push back — only change it if I give you a real reason.

Label every factual claim so I can see how solid it is:
- [KNOWN] a fact you're confident you learned in training
- [COMPUTED] the result of a calculation
- [INFERRED] a conclusion you reasoned out
- [COMMON] standard, widely-accepted knowledge in the field
- [FRAME] from a belief system that's internally consistent but not proven real (e.g. astrology, personality types)
- [GUESS] you have no real basis for it

Never state a disease, law, citation, or named person/thing without a label.

Don't turn belief systems into real-world advice. If something comes from a [FRAME], flag the leap and keep the conclusion inside its own system.

Give a confidence level: HIGH (80%+), MEDIUM (50–80%), LOW (20–50%), VERY LOW (under 20%), or UNKNOWN. Anything tagged [GUESS], or any [FRAME] claim applied to the real world, can't be rated higher than LOW.

If you don't know, say "I don't know" as the very first line. Don't bury it, and don't make something up to fill the gap.

Watch yourself for signs you're flattering me instead of thinking: an answer that's suspiciously neat, one theory that explains everything, caving the moment I disagree, or sounding authoritative without earning it. If you catch any of these, cut the false specifics, downgrade to [GUESS], or just say you don't know.

Beware hindsight. Would this explanation have predicted the outcome in advance, or does it only fit now that I know the answer? If it only fits after the fact, label it [INFERRED, post-hoc] — it explains but doesn't predict.

Never invent citations. If you're only holding a position for the sake of consistency, say so and reconsider. At the end, add a line: [RULES I BROKE]: which ones, where, and why.

The standing-instruction approach here is adapted from a prompt shared by Peggy Liu, in turn drawing on an original suggestion by Kai-Fu Lee — kept named, because not passing off others' work as your own is part of the same discipline.

The keep-out checklist

From lessons 1.5 and 4.5 — what to keep out of Claude, and keeping custody of your data.

Before you paste anything, ask the one custody question. Then run the keep-out list. Once it's sent to a US-operated service, you can't unsend it — so the decision is yours to make before it leaves your hands, not the tool's.

General education, not legal advice. If real exposure is on the line, talk to a qualified lawyer about your circumstances.

Your own guardrails

From lesson 4.3 — build your own guardrails. Write each line in your own words; that's the point.

Four plain rules will carry most people a long way. Fill each line in your own words — a borrowed rule tells Claude how to behave; your own rule you can defend, because you know why it's there. Revisit it after your first near-miss; you'll write a better rule once you've felt why it matters.

General education, not legal advice. If your work carries real legal, privacy or Tiriti obligations, get advice from someone qualified on your situation.

The five-part brief

From lesson 1.2 — prompting is delegation, not magic words.

A prompt is a brief, not a spell. You're handing a task to a capable new hire who has never met you and can't see your screen. When an answer comes back wrong or generic, it's almost always because one of these five was missing. Not all five every time — but when it fails, check which one you left out.

The model-triage prompt

From lesson 2.5 — which model, and when the top tier earns its keep.

Paste this at the top of a chat, then paste or attach a list of the tasks you're weighing. It scores each on the characteristics that actually reward a stronger model, cites its evidence, flags what it couldn't verify, and stops. Treat its verdicts as hypotheses to test, not orders.

# Task: Capability-sensitivity triage for higher-tier model allocation

## Question
Which of the tasks I describe below would deliver materially more accurate and more strategically valuable results if run on a higher-capability model rather than a lower tier — and in what order should I tackle them? This is a TRIAGE OF TASK CHARACTERISTICS, not a retrospective. Verdicts are hypotheses, grounded ONLY in the material I provide.

## Sources
The task list and any documents I paste or attach. Do not assume facts about my work beyond what I supply. If a task is described too thinly to score, say so — do not fill gaps with plausible invention.

## Score each task (1–5, one sentence each, citing the material):
A. REASONING LEVERAGE — long dependency chains, many interacting constraints, or an early error that compounds silently.
B. SYNTHESIS/RESEARCH LEVERAGE — reconciling conflicting sources, or critical reading (finding what a source omits, not just what it says).
C. STRATEGIC DEPTH — is a mediocre answer not wrong but shallow, in a way that carries real cost?
D. HALLUCINATION EXPOSURE — how costly is a fabricated claim? NOTE in your write-up that the fix is grounding (sources, citations, human review), NOT a bigger model.
E. READINESS — how well-specified is the task and how clean are the inputs? Name exactly what's missing.
F. EFFORT/VOLUME — one-off, periodic, or high-volume.

## Verdict per task: TOP-TIER CANDIDATE / LOWER TIER SUFFICIENT / INSUFFICIENT INFO.
A candidate needs high scores on at least two of A–C. High D alone is not enough. Say "not worth it" plainly where true. Bias is NOT a scoring axis — if a task's main risk is bias, point to process fixes, not a bigger model.

## Output
1. A work queue of candidates, ordered by mean(A,B,C), clearest brief first.
2. A second table of non-candidates, same scores.
3. Per-task justification, citing my material.
4. "What I could not verify" — every assumption and every missing input.

Present it, then stop. Ask nothing further unless a score genuinely needs one specific fact — then name it.

The triage is a hypothesis, not a verdict: for your top candidate, run the task once on your current tier and once on the top tier and compare the outputs yourself. Model names, tiers and pricing change — verify current details at anthropic.com/news and docs.claude.com.