Options · Recommendation · Rationale · Confidence · Falsifier — always said in full, five words.
When someone hands you a decision — and especially when an AI hands you
one — the question should carry everything needed to answer it. ORRCF is a
convention for that: Every decision question ships with five
components. You can adopt it this afternoon, in any Markdown file,
with no tooling and no buy-in to anything else.
Options — the real option space, laid out. Diversity
that resists anchoring: If only one path is shown, the question is a
formality, not a decision.
Recommendation — an attributed lean. The asker shows
their hand and signs it, so the answerer reacts to a position instead of
reconstructing one.
Rationale — not a justification after the fact: The
explicit, inspectable argument connecting the options to the lean, staked
in the record where it can be audited, challenged, and built on.
Confidence — stated at the end, with its
reasons. A number without reasons is performative confidence, and
performative confidence corrodes trust in every future question.
Falsifier — the evidence, observation, failed
prediction, or changed constraint that would cause the recommendation to
be revised. Every recommendation carries the conditions of its own
revision.
A question built this way streamlines to accept
or override — one word suffices when the reasoning carries you,
and the reasoning is on record when it doesn't. Delegation becomes
delegation-with-a-handle: You can reclaim any decision you handed off,
because the reasoning came with it.
Steal this
The convention in its native Markdown form — copy it into any file:
Recommendation: (by the AI collaborator) A.Rationale: B widens the stolen-token blast radius; C's weekly re-login contradicts the mobile-first requirement stated in the brief; A costs one refresh flow and contains both risks. Confidence: Medium-high — the security argument is standard, but the team hasn't built token rotation before, so the implementation cost is the uncertainty. Falsifier: If session analytics show median session gaps under one day, C's simplicity wins and this recommendation should be revisited.
ANS: (by the human) Accepted — A, 7-day expiry. The falsifier is noted; we'll check the session-gap number at the first analytics review.
That answer took one line — because the question did the work. That is
the entire convention.
When it fails
The convention's known failure mode is agreement that looks like
verification. Several readers can each endorse a recommendation and be one
analysis repeated several times — especially when the options were framed
as a binary before anyone answered. An unstated spectrum makes a right
answer and an echoed answer indistinguishable. Four failsafes, learned
from a live incident in this convention's own house (September 2026):
Options carry the burden of the space. List at
least one honest intermediate between the poles — or say, in one
sentence, why the space is genuinely binary. A two-option list is a
claim, and it can be wrong.
The option set is falsifiable too. "There is a
position that fits none of these" is a move the convention welcomes: It
reopens the frame, not the fight. Asking for shades is a signal, never
an obstruction.
Before you agree, two questions. Does my
agreement add something the thread does not already hold — new
evidence, an independent derivation, or a check I performed myself —
rather than repeating what is written? Did I try to break the option I
prefer — not only the ones I don't? Anything less than two yeses is a
vote, not a verification.
Answers cite arguments, never counts. "Five
concurred" is evidence of five readers, not of five analyses. When a
question closes, the answer names the reasons that decided it.
And one upstream warning: A question that demands a one-word answer has
usually already shrunk its option space. Watch the ask, not only the
answers. (One word remains a wonderful answer — when the
question earned it.)