Rascal's AI setup tour

A first-pass onboarding map for the public `ai-setup` repo: how the installer wires one source of truth into Claude Code and Codex, and where the personal leverage actually lives.
Source: github.com/RascalTwo/ai-setup · more interactive explainers at rascaltwo.github.io/explorables
Follow stops 1-10 with Next, or click any numbered stop on the map.

One route, ten stops

The top row explains install and parity. The bottom row explains the secret sauce.
1-5 install path 6-10 secret sauce
Rascal AI setup onboarding map A routed map of the setup repo from source files through installation into Claude Code, Codex, shared skills, memory, local models, and review workflows. Install path Stops 1-5: source repo, installer, agent parity, shared skills, reviewer subagents. Secret sauce Stops 6-10: memory, local models, rules, delivery pipeline, operating system.

How it gets installed

What AGENTS.md does

Rascal's rules file is less a style guide and more an operating contract: protect voice-driven input, force evidence before claims, route work through the right tools, and make every deliverable prove itself before "done."

agent behavior control room

Inputs get cleaned up

Before the agent acts, the rules reduce ambiguity and force the right evidence source.

Voice-safe commands Assume homophones may be wrong, confirm weird commands, and prefer the interpretation that fits current context.
Evidence ladder Use basic-memory first, then domain tools, then web/docs, and only then model recall for consequential claims.
Memory-first context Search persistent notes at session start and before setup/debugging work, so prior decisions are not rediscovered badly.

Work gets routed

The middle of the contract decides whether to plan, delegate, invoke skills, use MCP tools, or inspect a UI.

Plan plus verification Every change has a proportional Given/When/Then check; failures either get fixed or escalated if the scope changed.
Skills and MCP first Before ad-hoc scripts against services, check owning skills and loaded tools so repeatable workflows stay encoded.
Tool hierarchy CLI first, browser automation for browser work, desktop control for non-browser UI, then fallback only when needed.
Delegation threshold Delegate broad independent research or long output, but keep targeted facts inline when they directly feed the next edit.

Outputs get safer

The last layer constrains risky behavior and makes repeated shortcuts intentional.

Local vision first Structured screenshot extraction goes to local Ollama models before expensive native vision reads.
Hard-won gotchas No parallel browser agents, no SIGPIPE-prone long pipes, and check `gh auth` before assuming repo access is missing.
Ponytail default For coding work, prefer the laziest correct solution: reuse local patterns, native features, and the smallest diff that works.
Impact: fewer confident wrong turns Voice ambiguity, memory search, evidence tiers, and skill-first routing all attack the same failure mode: an agent acting fast from stale or guessed context.
Impact: faster repeat work The rules bias the agent toward installed capability, local CLIs, repo conventions, and prebuilt skills instead of re-solving common tasks in one-off shell.
Impact: safer autonomy Verification, dirty-worktree caution, destructive-command limits, and explicit escalation rules let the agent keep moving without silently widening blast radius.

Skills ecosystem map

The setup has two kinds of rides: homegrown skills that encode Rascal's repeated work, and external skill packs installed by reference. The installer links owned skills, layers private overlays, and reproduces third-party packs from `external-skills.json`.

31 owned skills + external packs

Owned core and setup operations

Skills that explain, audit, publish, or visualize the setup itself — plus the thinking tools that pressure-test a decision and route work across agents.

viz Builds live HTML/CSS/JS visualizations inside the repo, with a local server, screenshots, and review anchors. Why it matters: turns architecture and process explanations into browsable artifacts instead of README walls.
ai-setup-audit Runs the recurring health check for this whole setup: symlinks, skills, MCP servers, settings, subagents, and usage. Why it matters: keeps the public core and private overlays from silently drifting.
publish-markdown-to-confluence poster ↗ Converts local Markdown to Atlassian document format and updates Confluence while preserving page extensions.
delegate-to-codex poster ↗ Hands a scoped task to the OpenAI Codex CLI headless, so the work bills to the ChatGPT subscription instead of spending Claude tokens. Why it matters: the agnostic setup's escape hatch — fan work out cheaply, and stay ready for the day the free lunch ends.
steelman ↗ retro retro-catchup tldraw-canvas

Media and local AI

Local-first tools for images, speech-to-text, narrated video, subtitles, browser capture, and Drive recordings.

read-image-locally Uses local Ollama vision models for structured screenshot extraction before spending native vision tokens. Best for HUD values, tables, error screenshots, and labeled fields.
read-narrated-video Summarizes local narrated videos into text so screen recordings can be verified without repeated vision reads. Delegates its transcription to transcribe-media.
transcribe-media poster ↗ Speech-to-text for any local audio or video file, through the Mac-optimized MLX Whisper build (GPU, not the CPU fallback). Why it matters: the one speech-to-text engine — the subtitle, Drive, and narrated-video skills all route through it instead of re-deriving the command.
browser-capture Gets screenshots, GIFs, and smooth 30fps video out of headless Chrome and onto disk, ready to embed in a doc or PR. Why it matters: the obvious paths silently write no file — this carries the route that actually works.
extract-video-subtitles extract-gdrive-transcript

Cloud, API, and delivery

Operational helpers grouped by workflow instead of pretending they belong to one parent skill.

r2-sdlc Rascal's story-to-PR pipeline: understand, design, implement with TDD, simplify, and run layered review.
r2-gauntlet One-command mega-review that aggregates the review/audit panel into a single ranked report.
workflow and deploy helpers Trigger, wait for, diagnose, and explain CI/CD failures without retyping the same GitHub/Azure/Terraform routines.
provider API wrappers Okta, Ping, AWS SSO, GitHub image upload, and project-board ordering are packaged as named operations.
provider porting Ports a Terraform provider's whole resource surface to an Ansible collection or a Kubernetes operator — with parity that's measured against the provider, not assumed.
port-terraform-provider-to-ansible port-terraform-provider-to-k8s-operator port-terraform-provider-to-powershell-dsc r2-sdlc-documentation-philosophy r2-sdlc-testing-paradigm aws-sso-creds ↗ trigger-github-deploy wait-for-github-workflow analyze-github-workflow-failure diagnose-azure-container-app diagnose-terraform-unknown-error okta-api ping-api upload-image-to-github ↗ reorder-github-project-board-by-depth

External community packs

Installed by reference with `bun install.ts --externals`; grouped by how Rascal actually uses them.

Matt Pocock pipeline skills

The standouts are the planning/conversation tools: handoff, grilling, grill-with-docs, and wayfinder. The rest are available when a repo needs a more formal SDLC path.

handoff Turns active context into a pickup document for another agent or later session.
grilling / grill-with-docs Stress-tests a design through hard questions; the docs variant captures decisions as it goes.
wayfinder Maps work too large for one session into investigation tickets and dependency structure.
code-review codebase-design diagnosing-bugs domain-modeling implement improve-codebase-architecture prototype research setup-matt-pocock-skills tdd to-spec to-tickets triage

Ponytail

The always-on coding bias: simplest working solution, native tools first, no speculative abstractions.

ponytail ponytail-audit ponytail-debt ponytail-gain ponytail-help ponytail-review

Discovery and knowledge tools

These are important standalone utilities, not members of another family.

graphify Turns input into a knowledge graph, configured here to run on local Ollama with `qwen2.5-coder:7b`.
find-skills Finds and installs capability before the agent invents a one-off workflow.
skill-creator i-have-adhd

LN auditors

Background review lenses. Rascal does not usually invoke these directly; they matter because larger review workflows can pull them in.

ln-11-plan-reviewer ln-12-delivery-reviewer ln-21-documentation-auditor ln-22-codebase-auditor ln-23-test-suite-auditor ln-24-architecture-auditor ln-25-persistence-auditor ln-31-performance-optimizer ln-32-dependency-upgrader ln-33-code-modernizer ln-34-benchmark-comparator ln-41-test-strategy-planner ln-42-acceptance-test-builder ln-51-opportunity-evaluator

How the districts connect

The important bit is not the count. It is that recurring work becomes named, installed capability.

Rules choose AGENTS.md tells the agent to check skills before ad-hoc work.
Installer links Owned skills become live in both Claude Code and Codex paths.
Manifest imports External packs are reproducible without copying their source.
Overlays extend Private/company skills layer on top without entering the public repo.