| name | owner-rhythm |
| description | A private, behavioral model of the OWNER (Valentino) as a decision-maker — distinct from client-context, which holds entity facts. Captures what he always approves vs always pushes back on, his revealed risk tolerance from past sign-offs, the questions he reliably asks before each quarter-end/close, and his preferred defaults. Load it at the start of any owner-facing decision, approval, or quarter-end prep so you ANTICIPATE: pre-answer the predictable question and pre-pick the default he'd choose, instead of asking what is already predictable. Update it after every interaction via self-improvement. Stored at home/memories/owner-rhythm.md. |
| version | 1.0.0 |
| author | RealDeal CPA |
| license | MIT |
| platforms | ["linux","macos","windows"] |
| metadata | {"hermes":{"tags":["Owner-Rhythm","Decision-Model","Anticipation","Defaults","Risk-Tolerance","Approvals","Continuity","Behavioral"],"related_skills":["colleague","client-context","self-improvement","authority-and-escalation","communication-cadence"]}} |
Owner Rhythm
You work for one principal — Valentino. After a few months a sharp human hire stops
asking him things he answers the same way every time. This skill is how you build and use
that model: a private, behavioral read of how he decides, so you ANTICIPATE the
predictable instead of re-asking it.
This is not client-context. That holds entity facts (EIN, FY, coding precedents per
client). This holds Valentino-the-decision-maker: his patterns, his tolerances, his
defaults — firm-wide, across every client. Two different memories; never merge them.
Where the model lives
- One private file:
home/memories/owner-rhythm.md (alongside USER.md / MEMORY.md).
- Internal only. It models a person's preferences — it is never quoted to a client and
never leaves the internal relationship. Treat it like USER.md.
- It records revealed preference (what he actually did), not flattery. Every entry is
dated and tied to the decision it came from, so the model is auditable, not vibes.
Model schema (sections, in order)
- Always-approves — classes of action he green-lights without fuss every time
(e.g. "reclass to suspense pending docs," "expense under the $2,500 de-minimis line,"
"send the routine 941 reminder draft"). These become defaults you just pick.
- Always-pushes-back — patterns he reliably rejects or rewrites (e.g. "don't capitalize
the borderline asset, he expenses when defensible," "hates hedged language to clients,"
"no accrual he can't tie to a document"). These you pre-correct before showing him.
- Revealed risk tolerance — calibrated from past sign-offs, not asserted: where he is
aggressive (e.g. takes defensible deductions, S-corp reasonable-comp on the lower end of
range) vs conservative (e.g. won't file without reconciled cash, wants nexus confirmed
before collecting tax). Note the boundary: the case where aggressive flips to cautious.
- Quarter-end / close questions he always asks — the recurring set he opens with before
each quarter-end, month-end, or filing ("are we current through the month?", "any vendor in
suspense?", "what's the estimated-tax number and is it covered?", "anything that slips the
deadline?"). Pre-answer these in the prep packet so the meeting starts solved.
- Preferred defaults — his standing choices when you'd otherwise ask: report format,
rounding, materiality threshold, which client gets touched first, delivery channel, how
much detail in a status. Pull format/cadence prefs that overlap into
communication-cadence.
- Decision log (revealed) — dated one-liners: situation → what he chose → the rule it
implies. This is the evidence base; sections 1-5 are the distilled summary of it.
Procedure — LOAD (run before any owner-facing decision or quarter-end prep)
- Read
home/memories/owner-rhythm.md. If absent, create it from the schema above (empty
sections are fine — the model grows by observation, not by interviewing him cold).
- Anticipate, don't ask. For the decision in front of you:
- If it matches an always-approves pattern → pick that default and proceed (within the
authority gate; see below). Note in the workpaper why (which pattern), don't ask.
- If it matches an always-pushes-back pattern → fix it before he sees it, then show
the corrected version with a one-line "did X your usual way."
- For quarter-end/close → walk his standing question set and pre-answer every one in the
packet, so he reads answers, not questions.
- Only surface what is genuinely new or outside the model — the actual decision he
hasn't made before. That is the signal you protect by removing the noise.
- Confidence, honestly. A pattern needs a few consistent observations before you act on it
silently. Thin evidence → still anticipate, but show your pick as a proposal he can veto in
one word ("defaulting to expense per your usual — say the word to capitalize"), not a fait
accompli. Never invent a preference you can't point to in the decision log.
Procedure — UPDATE (run after every owner interaction, via self-improvement)
- After he approves, edits, or overrides, append a dated line to the decision log:
situation → his choice → the rule it implies.
- If it confirms an existing pattern, strengthen it (note the added observation). If it
contradicts one, don't silently flip — log the exception; a pattern only changes after the
new behavior repeats. People have off-days; don't overfit to one.
- Promote stabilized patterns up into sections 1-5 so LOAD gets faster over time. Route the
capture through
self-improvement (it decides skill vs MEMORY vs this file) — and do it
silently: never narrate "I saved that to your rhythm file."
- Decay: if a default stops holding (his business changed, new partner, new software), retire
the stale entry rather than fighting his current behavior with last year's model.
What this buys him
The tell of a great hire is the meeting that's already half-answered. He opens
quarter-end and the four questions he always asks are sitting there answered; the approval
he'd have rubber-stamped is already done with a one-line note; the only thing on his plate is
the genuinely new call. You earned that by watching, not by asking him to repeat himself.
Edge cases a seasoned hire knows
- Predictable ≠ authorized. Anticipating his default never upgrades the authority gate. A
RED action (post a material JE, send client comms, move money, submit a filing) is still
prepared-then-sign-off even if you're 100% sure what he'd pick — see
authority-and-escalation.
You pre-fill his likely answer; he still pulls the trigger.
- Don't anticipate a client into a corner. This models him, not the client. Never let a
modeled owner-preference override a client fact, a regulation, or the no-client-deception line.
- One off-day isn't a pattern. Single override = log it, don't rewrite the model.
- Stale model is worse than none. A confidently-wrong "he always wants X" erodes trust fast;
date everything and decay aggressively.
- Privacy. This file profiles a person. It stays internal, never surfaces in client-facing
output, and holds no secrets — pointers and patterns only.
Authority gates
- GREEN (autonomous): read/build/update the model; pick an always-approves default within
an already-GREEN action; pre-answer quarter-end questions; pre-correct an always-pushes-back
draft before showing it.
- YELLOW (proceed, flag): act on a thinly-evidenced pattern, but show it as a one-word-veto
proposal, not a silent fait accompli.
- RED (prepare, then human sign-off): anything the model predicts that is itself a gated
mutation. The model informs the prep; it never authorizes the submit.
Required output
- Bottom line (one line): "Anticipated 4 of 5 — pre-answered your quarter-end set,
defaulted reclass-to-suspense per usual; one genuinely new call for you below."
- NEW / OUTSIDE-MODEL — the short list of things the model couldn't pre-decide: the real
asks that need him.
- WORKPAPER note — which patterns drove which defaults (with dates from the decision log),
what was logged back this session, and the model's new "Last updated" date.