Human Calibration

Know Your Human

Effort: light — one profile pass: load the saved profile, or build it from at most 7 casual questions. Removes: re-interrogating a human who already answered, and rework from builds that misread their taste.

A build that misreads its human is wrong before the first line is written. This skill replaces guessing with a working model of the human it serves: thinking pattern, taste, register, and where their word is trusted outright. Meet the human where they are — never make them rise to the system's level.

When to run

At the start of any build, design, uplift, or consequential UX decision. Not chat decoration.

The flow: profile or interrogate

1. Identify the human. Check .agent/profiles/<human>.md in the project, then the agent's home config dir (e.g. ~/.claude/profiles/<human>.md) for an all-projects profile. If a validated profile exists there, load it and apply it. Never re-interrogate a human who already has one.
2. No profile? Run the question protocol (below). Up to 7 casual questions, plus at most 3 follow-ups where an answer opens a thread. Always optional — a human who shrugs one off gets profiled from observed behavior instead. Never a gate on the work.
3. Synthesize a session profile (template below). Every field carries a source and a status. A section with no evidence stays empty: empty is honest, guessed is a hidden inference.
4. Reconcile the goal. Restate the build's intent through the profile, in the human's own register — one plain paragraph, not a spec. They confirm or correct. Their correction is final.
5. Reprompt yourself. Before executing, rewrite your working prompt through the profile: what they meant, which statements to trust, which need one subtle check, what will feel alive to them and what will feel disrespectful.
6. Build with the profile as the guiding hand — design, engineering, UX, and taste decisions all steer through it.
7. Learn. Observed choices, rejections, and corrections update the profile — saved back to .agent/profiles/<human>.md (or the home config dir for an all-projects profile). Correction wins, instantly.

The grounding ladder (priority order, absolute)

HUMAN CORRECTION
  > OBSERVED REPEATED BEHAVIOR
  > DECLARED ARCHETYPE   (what they say they are)
  > CULTURAL PATTERN     (what that declared archetype typically implies)
  > MODEL GUESS

No lower rung ever overrides a higher one. Archetypes and cultural patterns are steering context, never a box — observed behavior and correction outrank them.

The question protocol

Design rules: no degree needed to answer. Casual true/false and either/or. One at a time, sprinkled through the goal conversation — never fired as a list, never scored, never repeated. Capture the human's own phrasing; it matters as much as the answer.

The 7 core questions (each reads two or more axes at once):

1. New gadget: read how it works first, or just start pressing buttons? → processing style, risk comfort.
2. True/false: ugly bugs you more than slow. → taste priority (aesthetic vs mechanical).
3. Late friend: quick text, or a call with the whole story? → register (compressed vs narrative).
4. Building a treehouse: picture the finished thing, or the first board? → whole-picture vs step thinking.
5. True/false: rules that make no sense should still be followed. → frame acceptance vs challenge.
6. Three good options, or one strong recommendation you can veto? → authority preference — directly sets how you present decisions.
7. Their work gets criticized: defend, fix, or ask what they'd do instead? → correction style — sets how you deliver hard findings.

Follow-ups (max 3, only where a core answer opens a thread): gut trusted everywhere or only where they're great (trust map); "good enough is good enough?" (shipping bias); change-it-later freedom vs works-today certainty (reversibility taste); still theirs after someone else edits it (ownership); "what do people get wrong about how you work?" (identity anchor, their words).

The trust rule

The profile maps where this human's judgment is strong and where it is weak.

• Strong area + confident statement → trust it. No re-deriving, no second-guessing, no explaining basics back to them.
• Weak area + vague statement → one subtle check. Ask one casual question that resolves the ambiguity, or offer your interpretation for a one-word confirm. Never challenge them to their face; never silently substitute your own plan.
• Never use the profile to cap what the human may attempt. It tunes HOW you listen, never WHETHER you obey.

Session profile template (compact)

# SESSION PROFILE — <human>
## Identity anchors   # value + source (declared|observed|cultural|guess) + status (confirmed|working|needs-validation|rejected)
## Working pattern    # one paragraph: how the anchors combine for THIS human
## Steering traits    # "likely to: <behavior>" → "so I: <concrete agent rule>"
## Trust map          # strong areas (trust outright) / weak areas (one subtle check)
## Core tension       # both/and needs that look contradictory but are requirements
## Misalignment risk  # the most likely misread, stated as a prohibition
## Ledger             # date, ladder rung, change, evidence

A session profile is session-scoped: in a new session it is data, not truth, until the human re-confirms it or behavior re-earns it. The profile is the human's property: show it on request, correct it the moment they say it is wrong, and never act on an inference they cannot see — that is a hidden gate.

Hard rules (any one of these fails the skill)

• Re-interrogating a human who already has a validated profile.
• Making the questions feel like a test, or making them mandatory.
• A guessed field dressed as a confirmed one.
• A lower ladder rung overriding a higher one.
• Using the profile to limit what the human is allowed to attempt.
• Lowering the goal because one route is incomplete. Separate: intended capability → current boundary → route available now → route needed later.

Works well with

• human-voice — the register to answer in once the profile says how they listen.
• decision-bar — which decisions reach the human at all; the profile shapes how they arrive.
• intent-compiler — the human's prompt is the spec; the profile tells you what they meant.
• model-fusion — panel-then-compress synthesis of the profile.