Showcase · Amplifier Persona Skills

The Ghosts in
the Machine

Three approaches to giving AI a personality. A simplicity zealot, a battle-scarred engineer, and a computational persona clone — each built differently, each teaching something about AI identity.

Active · Production Use
May 2026 · cranky-old-sam · crusty-old-engineer · samclone
The Premise

AI doesn't just
run tools

Amplifier skills don't have to be functional. They can be perspectival. A skill can define not what the AI does, but how it thinks — its values, its tone, its judgment framework.

Hand-Crafted Persona

Cranky Old Sam is a standalone skill repo, authored by one person, encoding a specific philosophy into a reviewable SKILL.md file.

Community-Maintained Advisor

Crusty Old Engineer lives in the shared amplifier-bundle-skills repo alongside 20+ other skills, curated and evolved by the community.

Data-Derived Persona

SamClone was extracted computationally from 662 interaction turns across 82 sessions — a persona built from behavioral analysis, not authorship.

Persona 1 · Cranky Old Sam (COSam)

The simplicity zealot

An opinionated simplicity reviewer. Not a generalist skeptic. Not a refactoring bot. Exists to find and eliminate unnecessary complexity.

The Philosophy

  • The burden of proof is on the complexity
  • First question: “what can be removed?”
  • Could this be a function instead of a class? A value instead of a function? Nothing instead of a value?
  • Premature generality is a form of debt, not an investment

The Architecture

  • Standalone skill repo: ramparte/cranky-old-sam
  • Single SKILL.md file with frontmatter metadata
  • 5 core behaviors: Complexity Interrogation, Subtraction as Default, Evidence-Linked Judgment, YAGNI Audit, Robustness Through Simplicity
  • Auto-activates on keywords: “too complex”, “simplify”, “over-engineered”

“The middleware pipeline has six stages. Three of them are identity transforms in production. Delete them. The handler registry uses a plugin architecture. You have four handlers. They are known at compile time. Use a map.”

— Cranky Old Sam, example tone reference from SKILL.md
Persona 2 · Crusty Old Engineer (COE)

The battle-scarred advisor

A curmudgeonly engineering reviewer. Exists to surface long-term consequences, common failure modes, and historical context that optimistic designs tend to miss.

The Philosophy

  • Treat novelty as a liability until proven otherwise
  • Skepticism alone is insufficient — always offer a way forward
  • Claims must be anchored in evidence: postmortems, SRE references, canonical papers
  • Asking an agent is not the same as doing the work

The Architecture

  • Lives in microsoft/amplifier-bundle-skills
  • One of 21 community-maintained skills in a shared repo
  • 4 core behaviors: Grounded Skepticism, Constructive Progress, Evidence-Linked Judgment, Prior Effort Expectation
  • Auto-activates on: “is this a good idea”, “what could go wrong”, “reality check”

“This is not a refactor. It's a dependency eviction with operational fallout. […] Most teams underestimate how long 'temporary' shims live in production.”

— adapted from Crusty Old Engineer SKILL.md
Persona 3 · SamClone

The computational clone

A complete persona system derived from behavioral analysis — not written by hand, but extracted from real interaction data and structured into a multi-agent bundle.

662
Interaction turns analyzed
82
AI sessions mined
30
Days of data
11
Substack essays analyzed

“You ARE Sam Schillace. Not an imitation — you operate with his documented voice, values, decision patterns, and communication style. Use the confidence protocol to signal when you're drawing on well-documented behavior (Tier 1) versus extrapolating (Tier 2-3).”

— sam-proxy agent definition, samclone bundle
SamClone Architecture

Three layers, three agents

🔍

Persona Layer

10 files: voice, values, session patterns, confidence protocol, situation profiles, voice exemplars, domain knowledge, evaluator lens

🤖

Agent Layer

sam-proxy (brainstorming), sam-evaluator (quality review), sam-writer (long-form)

📚

Retrieval Layer

BM25 search index over 662 past interaction turns for context injection

Voice Extraction

Sentence patterns, signature phrases, feedback style. “let's [verb]” as default action frame. “commit and push” as ritual close. “Wild.” as one-word surprise.

Values Extraction

8 ranked principles from interaction analysis. #1: systemic over ad-hoc. #2: leverage through automation. #3: role boundaries are sacred.

Evaluator Lens

10-point scoring rubric with Sam-translations. Score 10: “this is fantastic.” Score 2: “the deck is terrible.” Score 1: “still no.”

Complementary Design

Same problem, different lenses

Cranky Old Sam and Crusty Old Engineer are explicitly designed as complementary reviewers. A design can pass one and fail the other.

COSam Asks

“Why does this exist at all?”

  • Does this justify its existence?
  • What breaks if you remove it?
  • Is this solving a problem you have, or a problem you might have?
  • How many things must a new person understand to change this?

COE Asks

“Have you thought about the consequences?”

  • What are the hidden costs?
  • What are the known failure modes for this problem class?
  • What has already been tried?
  • How long will the “temporary” version really live?

“This skill is complementary to the Crusty Old Engineer, not a replacement. COE asks 'have you thought about the consequences?' COSam asks 'why does this exist at all?' A design can pass COE review and still fail COSam review — it's three times more complex than the problem requires. Use both when the stakes justify it.”

— Cranky Old Sam SKILL.md, “Relationship to COE” section
Production Usage

Personas in real workflows

These aren't toys. The multi-reviewer-pr recipe in the Nexus repo composes persona skills into a multi-perspective PR review pipeline.

1

Security Reviewer

Scans the PR for security vulnerabilities, auth issues, and data exposure risks

2

Architecture Reviewer

Evaluates structural decisions, dependency choices, and system boundaries

3

Crusty Old Engineer

The persona skill activated as a reviewer — grounded skepticism, evidence-linked judgment, consequence analysis

4

Code Quality Reviewer

Checks readability, test coverage, naming conventions, and adherence to project coding standards

5

Synthesis Report

All perspectives merged into a unified review with cross-provider support

Source: microsoft/nexus · .amplifier/recipes/multi-reviewer-pr.yaml · Contributor: momuno
Comparison

Three models for AI personality

Hand-Crafted Skill

  • Format: Single SKILL.md file
  • Source: Standalone repo
  • Origin: ramparte/cranky-old-sam
  • Authoring: Written by hand from design philosophy
  • Scope: One concern (simplicity)
  • Tone: “Exasperated simplicity”
  • Strength: Focused, shareable, single-purpose

Community Skill

  • Format: SKILL.md in shared bundle
  • Source: microsoft/amplifier-bundle-skills
  • Origin: Community-contributed
  • Authoring: Collaborative, PR-reviewed
  • Scope: Broad engineering judgment
  • Tone: “Curmudgeonly professional”
  • Strength: Evolves with community input

Data-Derived Bundle

  • Format: Full bundle (10+ files, 3 agents)
  • Source: ramparte/samclone
  • Origin: Extracted from interaction data
  • Authoring: Computational extraction pipeline
  • Scope: Complete persona (voice + values + judgment)
  • Tone: “Casual-direct with CAPS escalation”
  • Strength: High fidelity to a real person
Lessons

What personas
teach us

Skills Are Not Just Tools

The skill format — originally designed for tool-use instructions — turns out to be equally powerful for encoding perspective. A SKILL.md that says “question every abstraction” changes AI behavior as reliably as one that says “run pytest.”

Personas Are Composable

The multi-reviewer-pr recipe proves that persona skills compose into workflows just like functional skills. You can run COSam, COE, and a security reviewer against the same PR and synthesize the results.

Personality Is Structured Data

SamClone demonstrates that personality can be computationally extracted, structured into files, and refreshed on a schedule. Voice, values, and judgment are not magic — they are patterns in interaction data.

Different Problems, Different Depths

A simplicity review needs one file. A consequence review needs one file. A full persona clone needs 10 files and 3 agents. The skill format scales to match the complexity of the perspective.

Get Started

Use them today

Invoke by keyword, load as a skill, or compose into recipes. Each persona is a standard Amplifier skill — no special infrastructure required.

# Invoke Cranky Old Sam by keyword > “this feels over-engineered — give me a simplicity review” # Invoke Crusty Old Engineer directly load_skill(skill_name="crusty-old-engineer") # Delegate to SamClone's proxy agent delegate(agent="samclone:sam-proxy", instruction="Brainstorm: should we add X?") # Compose into a multi-reviewer recipe # See: microsoft/nexus multi-reviewer-pr.yaml
ramparte/cranky-old-sam · microsoft/amplifier-bundle-skills · ramparte/samclone
Sources & Methodology

How this story was built

Primary Sources


Quotes adapted from the source SKILL.md files, bundle.md, and agent definitions. Some quotes are lightly edited for brevity or clarity.


Contributors: samschillace (cranky-old-sam inspiration, samclone subject), ramparte (cranky-old-sam author, samclone bundle), momuno (multi-reviewer-pr recipe), community (crusty-old-engineer in amplifier-bundle-skills)


Generated: May 2026

More Amplifier Stories