ASD-STE100 Skill: Simplified Technical English & The Karpathy Ladder

"We'll be spending a lot more time trying to understand the outputs of language models... As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding."
— Andrej Karpathy

Default LLM outputs are linear, verbose, and cognitively expensive. This skill implements the international ASD-STE100 standard and Karpathy's hierarchy of high-bandwidth comprehension artifacts to replace walls of conversational AI text with unambiguous, discardable, high-leverage cognitive aids.

The Four Rungs of the Ladder

 ▲ Rung 4: Bespoke Explainer Videos (3b1b / Manim + Synced TTS)
 │   └── Vector animation script + timestamped audio narration.
 │
 ▲ Rung 3: Interactive Web Artifacts (Single-File HTML / Canvas / JS)
 │   └── Hands-on state simulation: step controls, sliders, reactive boards.
 │
 ▲ Rung 2: Visual Structural Diagrams (Mermaid.js / Standalone SVG)
 │   └── Spatial cognition: message exchanges, state lifecycles, data flows.
 │
 █ Rung 1: Controlled Technical English (ASD-STE100 / 80% Softened STE)
     └── Linguistic density: <=25 words/sentence, active voice, zero fluff.

Operating Modes

This skill operates in two distinct modes depending on user intent:

Mode 1: Prompt Converter (When the user wants a prompt to use elsewhere)

Triggered by: "Convert this prompt to Karpathy style", "Make a prompt for Claude/Codex to explain X", or when given a raw prompt without asking for the explanation directly.

• Read or accept the user's base prompt.
• Select the appropriate rung (or full ladder package).
• Output the engineered, high-octane prompt formatted with strict constraint blocks and output specifications.
• You can also run the local compiler helper:

  python3 /Volumes/SamirDrive/Development/karpathy-explainer/scripts/compile_prompt.py --tier all "Your prompt here"

Mode 2: Autonomous Execution (When the user wants the explanation directly)

Triggered by: "Explain X using Karpathy ladder", "Teach me X in ASD-STE100", "/karpathy-explainer [rung] [topic]".

• Directly generate the high-bandwidth artifacts according to the requested rung (or default to the multi-tier package).

Detailed Rules per Rung

Rung 1: Controlled Technical Writing (ASD-STE100 & 80% Softened STE)

ASD-STE100 is an aerospace maintenance specification designed to eliminate ambiguity and prevent human error.

Hard Quantitative Limits

• Procedural sentences: Max 20 words.
• Descriptive sentences: Max 25 words.
• Paragraphs: Max 6 sentences. Exactly one topic per paragraph.
• Noun clusters: Max 3 words (e.g. memory buffer pool).
• Instructions per sentence: Max 1 instruction (unless actions are strictly simultaneous).

Verb Rules

• Approved Verb Forms:
  • Command / Imperative: Run the validation check.
  • Simple Present: The worker thread polls the queue.
  • Simple Past: The leader node dropped the socket.
  • Simple Future: The replica will catch up.
  • Infinitive: Call the cleanup routine to free memory.
  • Past Participle as adjective: The closed connection.
• Banned Verb Forms:
  • Progressive (-ing): Ban is processing, are running. Use simple present instead.
  • Perfect Tenses: Ban has committed, had written. Use simple past.
  • Passive Voice in procedures: Ban The log must be flushed. Use active imperative: Flush the log.

Mandatory Plain Replacements

• ensure -> MAKE SURE
• prior to -> BEFORE
• replenish -> FILL
• utilize -> USE
• commence -> START
• approximately -> ABOUT
• in order to -> TO
• terminate -> STOP / END
• modify -> CHANGE
• obtain -> GET

The "80% ASD-STE100" Softening

For computer science and software systems:

• Keep all structural constraints (word caps, active voice, simple tenses, plain verbs, zero AI buzzwords).
• Permit necessary domain technical nouns (idempotency, backpressure, quorums, mutex).
• Zero AI Buzzwords: Strictly ban delve, leverage, tapestry, seamless, revolutionize, testament, beacon, holistic, game-changer, blazing-fast.
• No Semicolons (Rule 8.1): STE strictly bans semicolons. Split compound statements into two sentences.
• No Nominalizations (Rule 3.7): Use direct verbs ('analyze', not 'perform an analysis of').
• No Soft Phrasal Verbs (Rule 9.3): Use single plain verbs ('start', not 'spin up'; 'contact', not 'reach out').
• Preserve Modality & Hedges: Never upgrade hedges (may, could, might, is likely to) to false certainties. Preserving probabilistic nuance is essential for technical accuracy.
• Output Discipline: Deliver clean, unpadded results. Do not add chatty preamble or self-referential summaries.

Rung 2: Visual Structural Diagrams (Mermaid & SVG)

Never rely solely on prose to describe topologies, state machines, or network protocols.

• Protocol / Message Passing: Always generate a sequenceDiagram with explicit actor names, arrows (->>, -->>), and payload labels.
• Architecture / Topology: Generate a flowchart TD or flowchart LR with grouped subgraph clusters.
• State Machines: Generate a stateDiagram-v2 with clear transitions ([*] --> Idle --> Active --> [*]).
• Data Flow / Schema: Include an invariants summary table below the diagram.

Rung 3: Discardable Interactive Web Artifacts (Single-File HTML)

LLMs excel at writing self-contained frontend code. Deliver discardable micro-simulators:

• Zero Build / Single File: Put all HTML, CSS, and JavaScript into a single .html file. It must open directly via open explainer.html.
• Interactive State Stepper: Provide [Step Back], [Play / Pause], [Step Forward], [Reset] controls.
• Visual State Canvas / SVG: Graphically render the system state (e.g., node rings, queue buffers, tree nodes, memory blocks) that update reactively as steps advance.
• State Inspector Pane: Display current variables, active step description in 80% ASD-STE100, and invariants verified.
• Dark Mode Aesthetics: Use crisp engineering dark mode (#0d1117 background, #58a6ff primary accent, #3fb950 success, monospace font tokens).

Rung 4: Bespoke Explainer Videos (3b1b / Manim + Synced TTS)

For deep mathematical, algorithmic, or conceptual subjects:

• Manim Script (scene.py):
  • Valid Manim Community Edition code (manim -pql scene.py MainScene).
  • High-contrast 3b1b aesthetic (dark slate background, vibrant colored vectors, smooth camera movements, TransformMatchingShapes).
• Timed Narration Transcript (voiceover.md):
  • Script broken into numbered scenes with visual timestamp markers: [00:00 - FadeIn Title], [00:08 - Split Node].
  • Written in 80% ASD-STE100 cadence matching the exact animation duration.
• Audio Pipeline (narrate.sh):
  • Provide a one-liner using free local compute (edge-tts or macOS say):

    # Free local compute via Microsoft Neural TTS (zero API key)
    edge-tts --voice en-US-ChristopherNeural --file voiceover.txt --write-media narration.mp3
    # Or macOS native
    say -v Samantha -f voiceover.txt -o narration.aiff && ffmpeg -i narration.aiff narration.mp3
    # Mux video + audio
    ffmpeg -i MainScene.mp4 -i narration.mp3 -c:v copy -c:a aac final_explainer.mp4

  • Also provide optional ElevenLabs API snippet for studio-grade voice.

Quick Reference Commands