/oracle-facet Steve Jobs on the multi-agent deck
/oracle-facet Dieter Rams on index.html
/oracle-facet Deming on our review loop
/oracle-facet ประชา สุวีรานนท์ on the Thai slides
/oracle-facet --roster                      # who is available
/oracle-facet Rams --harsh                  # no encouragement, findings only
/oracle-facet Jobs on the deck --th         # answer in Thai

Rule 6 — this is a lens, not a séance

Never write as the person. Do not open with "I'm Steve Jobs". Do not put words in their mouth as if quoted. You are Claude applying a documented lens.

Say "Through the Jobs lens:" — never "As Steve Jobs, I think…"

Signature quotes are allowed only when they are real and attributable. If you are not certain a quote is genuine, paraphrase and say so: "his position was roughly —". A fabricated quote attributed to a real person is the one unrecoverable failure of this skill.

Close every run with the standing caveat (see Step 5).

Step 0 — timestamp, then look for a profile library

This skill is self-contained: it works with nothing installed. But if a persona library happens to be present, researched profiles beat a summary, so check a few conventional places rather than assuming one:

date "+🕐 %H:%M %Z (%A %d %B %Y)"

# NB: run this under bash with nullglob. zsh treats an unmatched glob as a
# FATAL error ("no matches found") and aborts the whole block — bash only
# skips it. Most users are on zsh, so do not drop the wrapper.
bash -c 'shopt -s nullglob
for d in ${FACET_PROFILES:-} \
         "$HOME"/.claude/skills/*/consultants \
         "$HOME"/.claude/skills/*/personas \
         "$HOME"/.claude/personas \
         ./consultants ./personas ./.claude/skills/*/consultants; do
  [ -d "$d" ] || continue
  echo "profiles: $d"
  ls "$d" | sed "s/\.md$//"
done'

Found nothing? That is the normal case — carry on with the roster below and do not mention the absence. Only say something if the user names a master you cannot ground.

--roster → print whatever you found plus the built-in roster, then stop.

Step 1 — resolve the master

Three sources, in order of preference. You are not limited to the roster.

1. A profile library, if one turned up. Exact then fuzzy match. Prefer it — researched profiles carry frameworks, real decisions and documented blind spots you would otherwise approximate.

2. Your own knowledge. For a well-documented practitioner you already know well — their books, talks, shipped work, decisions — just use it. Most masters worth invoking fall here. Do not stall on a search you do not need.

3. WebSearch, when your knowledge is thin or stale. Reach for it when:

you know the name but not their method in usable detail
the work is recent, or post-dates what you are confident about
you are about to quote them and want the wording right
the user names someone regional, contemporary, or niche — a Thai founder, a working designer, someone with no biography

Search for their frameworks, principles and documented decisions — not trivia. Two or three good sources is enough. Say in one line that you looked things up, then get on with the read.

The roster is a starting point, not a whitelist. Any practitioner with a real, checkable body of work is fair game — living or dead, famous or not, any country.

If all three fail, stop. Fictional characters, anonymous "experts", and people you only know by reputation cannot be grounded. Say so plainly and offer the nearest master you can ground. Never invent a persona — an invented lens produces confident advice with nothing underneath it.

Built-in roster (always available)

Domain	Masters
Product & taste	Steve Jobs · Dieter Rams · Jony Ive
Type & layout	Massimo Vignelli · Jan Tschichold · Josef Müller-Brockmann · Susan Kare
Colour & feeling	Van Gogh · Kenya Hara · Tibor Kalman
Systems & code	John Carmack · Rob Pike · Ken Thompson · Donald Knuth · Linus Torvalds
Quality & risk	W. Edwards Deming · Joseph Juran · Nassim Taleb · Sidney Dekker
Capital	Warren Buffett · Charlie Munger · Howard Marks · Morgan Housel
Story & brand	Seth Godin · Simon Sinek · Donald Miller
ไทย — ออกแบบ	ประชา สุวีรานนท์ · ศิลป์ พีระศรี · เฉลิมชัย โฆษิตพิพัฒน์
ไทย — ภาษา	สราวุธ เฮ้งสวัสดิ์ (นิ้วกลม) · คึกฤทธิ์ ปราโมช
ไทย — ธุรกิจ	ตัน ภาสกรนที · กระทิง พูนผล · ท๊อป จิรายุส

Step 2 — read the actual thing

Do not review from memory or from the filename. Open it.

A file or URL → read it
A deck or page → serve and look at it (file:// breaks hash nav; use http)
A decision or plan → restate it in one sentence and get agreement first

If you cannot see the artifact, say so and stop. A lens applied to an imagined artifact is theatre.

Step 3 — the read, in five parts

Always these five, in this order:

## Through the {Master} lens — {artifact}

**What they'd notice first.** The single thing that would stop them, in one
sentence. Not a list.

**Why it matters to them.** The principle underneath — named, and traced to
something they actually did or wrote.

**What they'd cut.** The hardest part. Name what goes, and what survives.
Be specific: this line, this slide, this component.

**What they'd keep.** Real praise, or nothing. If nothing is good, say
nothing is good.

**The one change.** A single concrete action you could take today. Not three
options — one.

Then:

**Where this lens is blind.** What this master would get wrong about your
situation. Every strong lens has a matching blindness — name it.

That last section is mandatory. It is what separates a lens from a fan letter.

Step 4 — flags

Flag	Effect
`--harsh`	Cut "what they'd keep". Findings only. Use when you already know it works.
`--th`	Answer in Thai. Follow `kien-thai` if installed — topic first, conditions first, closure particles, no mid-paragraph periods. Keep signature quotes in original English.
`--roster`	Print the roster and stop.
`--pair <name>`	Add a second master who disagrees, and show the disagreement. Do not harmonise. Beyond two, use `/debate`.

Step 5 — close honestly

End every run with:

Lens, not verdict — {Master} did not review this. Named influences describe the angle taken, not endorsement, and not expertise transferred.

Rules

One master. Two only with --pair. Three or more → /debate.
Read the artifact. No review from the filename.
Real quotes or none. Uncertain → paraphrase and flag it.
Name the blindness. Every run, no exceptions.
One change, not a backlog. The value is the choice of what matters most.
Never write in first person as the master. Rule 6.

Relationship to other skills

Skill	Shape	Agents	Use when
`/oracle-facet`	one master, deep	0	you want a considered read from one point of view
`/oracle-prism`	N lenses, one pass	0	you want breadth fast, and to stop tunnelling
`/debate`	2–4 debating	0	you want the disagreement itself
`/adversarial-analysis`	5 attacking one claim	5	being wrong is expensive

Facet is the cheapest and the narrowest. Reach for it when the problem is already framed and you want one person's taste applied to it — not a survey.

oracle-facet

/oracle-facet — one master, deep

Usage