The Ghost You Can Review

A personality, extracted from real interactions

Amplifier · Persona Skills

The Idea & The Catch

A persona skill encodes how an AI thinks — but who says it's faithful?

An Amplifier skill can capture not just what an AI does but how it thinks — a real point of view. Hand-write that persona, though, and it's only one author's opinion, and no one can prove it's true to the person.

So the question becomes: could a point of view be extracted from evidence instead of asserted?

The Proof

SamClone didn't invent a personality — it extracted one from real interaction data

The extractor scanned root Amplifier sessions over a 30-day window and produced a corpus of real turns — not a prompt someone typed from memory.

Raw evidence answers "just one person's opinion." Now it has to become reviewable.

662

interaction turns

82

conversation sessions (of 90)

15

projects spanned

The Artifact

The corpus becomes a layered persona — nine files, not one monolithic prompt

SamClone distills the raw turns into focused, reviewable context files in the bundle — each one small enough to actually read and check.

Splitting it up is what makes the persona auditable — starting with how it thinks.

voice.md — 56 lines
values.md — 56 lines
session-patterns.md — 54 lines
confidence-protocol.md — 73 lines
situation-profiles.md — 91 lines
+ voice-exemplars, writing-voice, domain-knowledge, current-state

How It Thinks

The values file makes judgment explicit — ranked principles and named mental models

This is the part that captures thinking, not just phrasing: 8 ranked core principles and 6 named mental models the persona reasons with.

But if a clone reasons like this, a hard objection follows.

The Radius
The Curve, Not the Point
Taste as Scarce Resource
Machine-That-Builds-Machines
Human Attention as Bottleneck · Software Factory Commoditization

The Killer Objection

How do you know the clone isn't just confidently making things up?

An extracted persona invites the sharpest challenge of all — fidelity. A convincing voice that fabricates is worse than no clone at all.

This is the moment the whole idea either survives or collapses.

The Turn

A 3-tier confidence protocol makes fidelity auditable instead of bluffing

The persona is forced to signal how it knows what it knows — and a meta-rule says genuine uncertainty in Sam's voice beats confident extrapolation.

Reviewability isn't a slogan now — it's an enforced rule. So it can go to work.

Tier 1 · Documented — behavior grounded in real captured turns
Tier 2 · Interpolated — reasoned between documented points
Tier 3 · Extrapolated — flagged as beyond the evidence

The Composed Ghost

The extracted ghost becomes a working teammate that reasons, critiques, and writes

A real local Okapi BM25 index (1,294,509 bytes, no external API) retrieves relevant past turns and injects them into three specialized agents.

The persona isn't a document anymore — it's a composable, live workflow.

sam-proxy

model role: reasoning

sam-evaluator

model role: critique

sam-writer

model role: writing

The Pattern To Keep

A whole point of view lives as ~3,637 reviewable lines that refresh weekly

Voice, judgment, and an evaluator's 1–10 quality bar were committed as v0.2.0 — and a weekly pipeline regenerates current-state.md from the last 7 days.

Personality stopped being magic. It's now something you can extract, audit, and engineer.

3,637

lines across 19 files (v0.2.0)

7

day refresh window

1–10

evaluator quality rubric

Sources

Sources & Research Methodology

Data as of: extraction run 2026-03-14; session activity 2026-03-03 to 2026-03-14 (30-day scan window).

Feature status: Working & reproducible — bundle v0.2.0; 2 commits (ada590f, dfc5275).

Primary source: ramparte/samclone. Author: Sam Schillace; Co-Authored-By: Amplifier.

Research performed:

Turn count: wc -l extraction/raw/all_turns.jsonl = 662; session_summary.json total_turns=662
Sessions/projects: cat extraction/raw/session_summary.json + python aggregate = 82 of 90 sessions, 15 projects
Persona files: wc -l persona/*.md persona/*.yaml; bundle.md context include = 9 files
Judgment model: cat persona/values.md = 8 ranked principles, 6 mental models
Fidelity: cat persona/confidence-protocol.md = 3 tiers + meta-rule
Retrieval: ls -la retrieval/index.pkl = 1,294,509 bytes; python3 retrieval/get_context.py returned live ranked results
Agents: grep model_role agents/*.md = sam-proxy/reasoning, sam-evaluator/critique, sam-writer/writing
Refresh: wc -l refresh/update_current_state.py = 591; DEFAULT_WINDOW_DAYS=7
Scale: git show --stat ada590f = 19 files changed, 3637 insertions(+)

Gaps & notes: The "30 days" is the scan window; captured data spans ~11–12 days. bundle.md rounds the 66 tagged operational quotes to "~60." The git author email sam@example.com is a placeholder. In-repo biographical claims about Sam are the repo's own assertions, not independently verified.