Sethu Nguna.
Independent R&D · 2026

Neckbeard

Agent skill

An agent skill that reads a codebase and drafts the learning around it: onboarding, task tutorials, practice and recovery. Every claim is marked by what backs it: what the agent ran, what it only read, and what is still unknown.

← Explore all projects
Audiences
End users, administrators, developers, support teams and people adopting a new release
Runs in
Codex, Claude and Cursor, through portable instruction adapters
Components
A required core plus eight optional components, from end-user onboarding to media and release training
Built with
Agent instructions and Python standard-library scripts for install, governance, drift checks and media gates
Overview, 2:20, with captions: the README says one thing, the code decides another, and Neckbeard reads the code.
The design problem

The challenge

A README describes what software should do, and the code decides what it actually does. Learning material written from the README, by people or by AI, can confidently teach steps nobody has checked, and it drifts as the code changes.

The logic

Treat learning content the way engineers treat code. Separate what was run from what was only read, tie every document to the exact source revision it describes, and flag a guide as stale when the code it depends on changes.

What it produces

Onboarding paths, first-task tutorials with real commands and recovery steps, practice, and a governance record that can be exported as an audit package.

My role

I designed and built Neckbeard as independent R&D.

Worked case study

Fractured Order on Plenum

The repository includes a case study that teaches a Blue Scribe’s first Strategic Orientation handoff in a specific Plenum deployment. It comes with a two-page PDF guide, a job aid, a narrated explainer video, audio, captions, a transcript and a transfer exercise.

What it shows, and doesn’t

A synthetic rehearsal of the application passed. How it behaves on the hosted platform, and what learners take from it, remain unverified.

How a run works

Read, verify, teach, govern.

The same sequence whichever audience the learning is for.

01

Load the profile

An organisation profile sets house style, terminology, approved tools, classification rules and reviewers. Without one, outputs stay drafts.

02

Inspect the repository

The agent reads the implementation and keeps runtime checks separate from code inspection.

03

Draft the learning

Orientation, task tutorials, practice and recovery, using whichever components the audience needs.

04

Record the evidence

Every document carries an owner, the exact source revision, its evidence state, a reviewer and a review-by date.

05

Gate the media

Narration, screenshots and video only go through approved tools, modes and data classifications.

06

Watch for drift

A drift check compares code changes to the paths each guide depends on, and flags guides that may be stale.

Design decisions

Every rule has a job.

Neckbeard is built so that a guide can’t claim more than its evidence supports.

DecisionWhat it protectsIn the skill
Ran versus read Trust in what a tutorial says will happen. Evidence states separate what the agent ran from what it only inspected.
No claims without learners Honesty about effectiveness. The skill doesn’t certify usability or learning effectiveness without learner evidence.
Drafts until configured The organisation’s control over what ships. With no configured profile, outputs stay drafts and external media processing is blocked.
Every document traceable Reviews and audits. Owner, source revision, evidence state, reviewer, review-by date and a stable ID on every output.
Approvals bound to the exact files Release readiness that means something. Each approval records a hash of the approved files and the full source snapshot.
Drift caught in CI Guides that stay current. A workflow checks pull requests, and the drift check flags guides whose source has changed.
Modular components Learning that fits the audience. Install only what’s needed: end-user, admin, developer, practice, maintenance, media, support or release training.
Evidence

What’s documented

The repository includes runnable product and service fixtures, five behavioural evaluation cases with a weighted scoring rubric, automated tests of the governance and distribution scripts, and the worked Fractured Order case study.

Limits

What the evidence can’t show

Live behaviour in each host agent hasn’t been certified, and passing code tests say nothing about learning outcomes. The case study’s hosted behaviour and learner results are unverified.

Supporting artifacts

Explore the work.

Repository ↗

The skill, its components, governance scripts, evaluations and the Fractured Order case study

Start a conversation

Build the capability.
Make it work in practice.

Learning strategy, institutional systems and decision practice.

sethnguna@gmail.com ↗