AI Guides › Claude Mastery
Claude For Decisions: Borrowing The Board's Evidence Ledger For Your Own Chats
By Nigel Guy · 2 min read
Most people ask Claude for a decision the same way they'd ask a friend at a
bar — "what do you think I should do?" — and get back a plausible-sounding
answer built on whatever they happened to mention. The problem isn't the
answer's quality. It's that nobody separated what was actually known from
what was just assumed before the advice got generated.
The rule: label your inputs before you ask for a recommendation, and the
recommendation gets dramatically more useful — you don't need our full
advisory-board product to do this, just its opening move.
The mechanism
Before you ask "what should I do," spend two minutes sorting what you know
into four labels:
- Fact — something you've actually verified or observed.
- Assumption — something you believe but haven't checked. (Most of what
people call "facts" in a first message are actually this.)
- Unknown — something material you genuinely don't know yet.
- Constraint — a hard boundary the answer has to respect (budget, time,
a non-negotiable).
Then:
- Paste the labelled list in before you ask the question. Not "should I
raise my price?" — "here's what I know (labelled), here's what I'm
assuming, here's what I don't know yet, here's what's fixed — now, given
that, what should I do?"
- Ask it to flag which of your assumptions the recommendation actually
depends on. This is the single highest-value follow-up question in this
whole guide.
- Ask what would change the answer. A recommendation that can't name
the thing that would flip it is closer to a guess with good formatting.
- Get a stop list, not just a to-do list. A useful recommendation says
what to not do as clearly as what to do.
What to skip
Skip building an elaborate spreadsheet for this — it's a two-minute mental
sort, not a project. And skip it entirely for genuinely low-stakes calls
(what to have for lunch doesn't need an evidence ledger); save the discipline
for decisions that actually cost something to get wrong.
Guardrails
- This is a lightweight version of a real practice, not a substitute for it —
bigger, genuinely risky decisions deserve the full process (multiple
perspectives, a real test, a written threshold), not just a labelled list
and one chat.
- Labelling something a "fact" doesn't make it one — the discipline only
works if you're honest that most of what feels certain is actually an
assumption wearing a confident sentence.
- A confident recommendation built on your own mislabelled assumptions is
still built on mislabelled assumptions. The ledger reduces this risk; it
doesn't eliminate it.
All 751 AI guides · JulieMango plans from £17/mo