AI Experiment 01: The Chord Tone Atlas

THE MAP IN BRIEF.

The Problem: My improvisation sounds "scotch-taped"—disconnected licks pasted over a chord progression rather than lines that follow the harmony.

The Contrarian Insight: Most guitar pedagogy teaches you to think in shapes (pentatonic boxes, CAGED patterns, modal positions). This is backwards. The pros think in chord tones—the 1, 3, and 5 of each chord—and use voice leading to connect them. Shapes are a side effect of this deeper logic, not its foundation.

The LLM Leverage: Mapping chord tones across a 12-bar progression is tedious grunt work. Claude excels at grunt work. I provide the progression; Claude generates a visual reference card (the "Atlas") showing chord tones, roman numeral analysis, and voice leading connections. The artifact fits on a 4x6 notecard. I practice with it for five days. I film the results. I repeat with the next song.

The System: A templated, metadata-rich notecard format that grows into a searchable personal deck of progressions over time—organized by genre, difficulty, harmonic feature, or technique focus.

The Outcome: Lines that lock into the changes. A growing library of harmonic maps. A reusable framework for any chord progression I encounter.

THE CONTEXT.

Perhaps a strange first experiment for a blog ostensibly about "using LLMs to break through guitar plateaus." No prompt engineering wizardry. No generative AI composing licks for me. No fancy Markov chains or Bayesian inference applied to fretboard navigation. Just... a reference card.

But that's precisely the point.

In my pursuit of fluent improvisation—the kind where your lines actually follow the chords rather than float generically above them—I kept encountering the same wall. I know my scales. I know my modes. I can shred through a B minor pentatonic box at tempo. And yet, when I solo over a progression, something's missing. The lines sound assembled, not composed. Scotch-taped. As if I'm playing at the music rather than with it.

The blinding flash of the obvious arrived recently: I don't actually know where the chord tones are.

Not abstractly—I understand that Bm contains B, D, and F#. But locationally, in real-time, across the fretboard, while the changes are flying by? That's a different animal. I've been improvising from shapes (visual patterns memorized by rote) rather than from harmony (the actual notes that define each chord). The shapes are a crutch that obscure the underlying logic.

This is, I suspect, a common plateau. And it's one where LLMs can provide genuine leverage—not by generating music, but by doing the tedious cartographic work that makes practice more targeted.

THE PROBLEM, STATED PRECISELY.

01 | The Shape Trap.

Most guitarists learn improvisation through visual patterns: the five pentatonic boxes, the CAGED system, the seven modal shapes. These are useful scaffolds—but they're also seductive. Once you've memorized them, you can "improvise" by running up and down familiar shapes without ever engaging with the harmony of the progression underneath.

The result? Generic lines. Licks that work over any minor chord because they're not specific to this chord. The dreaded "noodling."

02 | The Chord Tone Gap.

Jazz players and session pros approach improvisation differently. They think in chord tones—the root, third, and fifth of each chord—and use these as anchor points. Everything else (scale runs, chromatic approaches, tensions) is ornamentation around these structural pillars.

The problem: mapping chord tones across a full progression requires tedious work. For a 12-bar tune with 6-8 distinct chords, you're looking at 18-24 notes to locate across the fretboard, plus understanding how they connect (voice leading). Most players skip this step and default back to shapes.

03 | The Voice Leading Void.

Even if you know your chord tones, connecting them smoothly across chord changes is a separate skill. Voice leading means finding notes that belong to both chords and using them as bridges. For example: when moving from Bm (B-D-F#) to Dmaj (D-F#-A), the notes D and F# are shared. Landing on one of these at the moment of the chord change creates smooth, connected phrasing.

This is what separates "lines that follow the changes" from "scale runs over a backing track."

THE EXPERIMENT.

Hypothesis: If I use Claude to generate a visual reference card—a "Chord Tone Atlas"—for a specific progression, showing (a) chord tones with intervals, (b) roman numeral analysis, and (c) voice leading connections, then practicing with this artifact will accelerate my ability to improvise fluently over the changes.

Test Case: Jason Becker's "Valley of Fire" (from Triumphant Hearts, 2018). A 12-bar progression in B minor:

The Prompt:

What Claude Returns: A structured HTML card with:

• Metadata header (song, artist, key, tempo, difficulty, practice tracking)

• Chord grid showing tones + intervals + voice leading for each bar

• Warning flags for non-diatonic chords (F#maj uses A# instead of A natural—outside B minor)

• 5-day workout block with progressive exercises

The key design decision: no fret positions on the card itself. The Atlas shows what to play (note names, intervals). The where (fret positions) becomes the constraint built into the workout protocol. This separation keeps the reference card clean and forces me to learn the fretboard rather than memorize another shape.

THE ARTIFACTS.

I've constructed three artifacts from this experiment:

01 | The Completed Example.

A fully worked Atlas card for "Valley of Fire" showing:

• Chord tones for all 10 distinct chord positions in the 12-bar loop

• Roman numeral analysis from D Major: vi → I → III → vi → ii → V → vi → vi → ii → IV → III → III

• Voice leading instructions (e.g., "Hold D or F# into next chord")

• Warning about the borrowed III chord (F#maj contains A#, not A—this is outside the B natural minor scale)

• Block 1 workout: "Chord Tone Geography" (5 days, 20 minutes each)

02 | The Blank Template.

A reusable HTML template with:

• All styling pre-built

• [BRACKETS] indicating fields to fill in

• Consistent visual design for deck coherence

• Metadata schema for searchability

03 | The Metadata Guide.

Documentation explaining:

• What each metadata field means and why it matters

• File naming conventions (YYYYMMDD_SongTitle_Artist.html)

• Deck organization strategies (by genre, difficulty, harmonic feature, technique)

• Workflow for creating new cards

Together, these artifacts constitute a system—not just a one-off exercise, but a repeatable framework for any progression I want to internalize.

THE ROMAN NUMERAL ANALYSIS.

From the perspective of D Major (treating B minor as the relative minor):

The F#maj Problem.

Bars 3, 11, and 12 feature F#maj—a chord that doesn't belong to the diatonic D major / B natural minor scale. In diatonic harmony, this chord would be F#minor (iii). By making it major, Becker introduces A# (the major third of F#) instead of A natural.

This matters for improvisation: if you're running B natural minor patterns over these bars, you'll clash. The A# creates a leading tone that wants to resolve up to B—which is exactly what happens when F#maj moves back to Bm. This is modal interchange (borrowing from the parallel minor's dominant area), and it's the harmonic "spice" of the progression.

The Atlas flags this so you don't get blindsided.

THE WORKOUT: BLOCK 1.

Theme: Chord Tone Geography.

The goal of Block 1 is simple: learn where the chord tones live on the neck. No improvising yet. No creativity. Just cartography.

Day 1 | One String Hunt (20 mins)

• Pick ONE string (high e or B recommended)

• Find all 3 tones for each chord on that string only

• No backing track, no metronome

• Go slow. Speak the note names out loud as you play them

• Target the unusual chord first: F#maj (F#-A#-C#)

Day 2 | Two String Combinations (20 mins)

• Pick TWO adjacent strings (e.g., B + high e)

• Find all 3 tones across both strings

• Note: which string pairs feel natural? Which feel awkward?

• Still no backing track—just geography

Day 3 | Position Clusters (20 mins)

• Pick a fret region (e.g., 7th-10th fret area)

• Find all 3 tones for EACH chord within that box

• Don't move your hand position—stay in the zone

• Repeat in the 12th-15th fret region

• Question to answer: which chords cluster together? Which require stretches?

Day 4 | Chord Sequence, No Improv (20 mins)

• Loop backing track at HALF SPEED (~56 BPM)

• Play exactly 3 notes per bar: root, 3rd, 5th

• Any order, any string, any position—just hit the tones

• Goal: make the changes. No fills, no scale runs

• If you miss a change, keep going—don't restart

Day 5 | Diagnostic Recording (20 mins)

• Full tempo backing track

• Improvise using ONLY the 3 chord tones per bar

• RECORD this session

• After: listen back and mark:

  • Which chords felt blind?

  • Which positions did you gravitate toward?

  • Did you hear the A# in F#maj, or did you play A natural by mistake?

Block 1 is not glamorous. It's not creative. It's the equivalent of a jazz student doing "chord tone only" choruses over a standard—except I'm doing it over Becker's neoclassical shred progression. The point is to build the foundation before adding ornamentation.

WHY LLM, EXACTLY?

A reasonable question: couldn't I do this with pencil and paper? Why involve Claude at all?

The honest answer: I wouldn't do it otherwise.

Mapping chord tones, calculating voice leading, formatting a printable reference card—this is 30-60 minutes of tedious work per song. The cognitive overhead is just high enough that I'd skip it and default back to scale shapes. The LLM removes the friction.

More importantly, the LLM enables systematization:

• Consistent formatting across all cards (same visual hierarchy, same metadata structure)

• Searchable metadata for future retrieval (find all level-3 jazz progressions with secondary dominants)

• Reusable template for rapid card generation

• Evolving deck that grows with my practice

This is the "via negativa" application of LLMs to guitar learning: not adding AI-generated licks or flashy generative features, but subtracting the grunt work that stands between me and focused practice.

The map != the terrain. But you still need maps.

THE GENERALIZATION.

This framework works for any chord progression in any style:

1. Identify the progression (by ear, from a chart, or by asking Claude to help transcribe)

2. Run the prompt (specifying key center, tempo, genre, difficulty)

3. Generate the Atlas card (chord tones, roman numerals, voice leading, warnings)

4. Complete Block 1 (Chord Tone Geography, 5 days)

5. Assess and iterate (record Day 5, identify weak spots, design Block 2)

Future blocks might focus on:

• Voice Leading Chains (connecting chord tones across changes)

• Rhythmic Phrasing (when to play vs. when to rest)

• Approach Notes (chromatic and diatonic approaches to chord tones)

• Dynamic Control (volume swells, ghost notes, accents)

But I won't pre-engineer these. The beauty of the thematic block structure is that it adapts to what I actually need after completing the previous block. Over-structuring is death in an improvisational context.

THE DECK, OVER TIME.

If I create one Atlas card per week, I'll have 50+ cards within a year. At that scale, the deck becomes a knowledge base:

• Filter by genre: Show me all jazz standards I've mapped

• Filter by harmonic feature: Show me progressions with borrowed chords

• Filter by difficulty: What are my level-2 cards for warm-up days?

• Filter by technique: Which cards emphasize wide position shifts?

The metadata schema makes this possible. Each card is tagged with:

This isn't just practice material—it's an evolving personal curriculum, shaped by the progressions I actually care about rather than some generic syllabus.

THE DEMONSTRATION.

[Vimeo embed: Before/after comparison over Valley of Fire backing track]

• Before (Day 0): Generic B minor pentatonic runs, no connection to chord changes, occasional clashing over F#maj bars

• After (Day 5): Lines targeting chord tones, smooth voice leading across changes, conscious acknowledgment of A# in F#maj bars

The difference isn't virtuosic. It's structural. The lines sound like they belong to the progression rather than floating generically above it.

WHAT THIS IS NOT.

This experiment is not:

• AI-generated music. Claude doesn't compose licks for me. I compose them, informed by the harmonic map.

• A replacement for ear training. The Atlas is a scaffold, not a crutch. The goal is to internalize the information until I don't need the card.

• A comprehensive method. This is one tool among many. It addresses a specific plateau (chord tone fluency), not the entirety of guitar mastery.

• A claim that LLMs "understand" music. Claude is a text prediction engine. It generates useful maps because chord-tone-to-note-name mapping is a well-defined, information-processing task—not because it "hears" harmony.

What this experiment is: a demonstration that LLMs can provide genuine leverage for guitar practice when applied to the right problem (tedious information work) with the right artifact output (a reusable, physical reference).

THE FILES.

Available for download:

1. VOF-001_ValleyOfFire.html — Completed example card

2. chord_tone_atlas_TEMPLATE.html — Blank template for new cards

3. ATLAS_METADATA_GUIDE.md — Full documentation of the metadata schema

4. README.md — Quick-start guide and deck organization strategies

NEXT EXPERIMENTS.

Ideas percolating for future String37 entries:

• Experiment 02: Tone Archaeology — Using Claude to reverse-engineer Quad Cortex patches from reference recordings

• Experiment 03: The Rhythmic Displacement Engine — LLM-generated polyrhythmic exercises

• Experiment 04: Interval Prohibition Drill — Banning your crutch intervals to force vocabulary expansion

• Experiment 05: Position Genetics — Evolutionary mutation across fretboard positions

Each experiment will follow the same structure: Problem → LLM Leverage → Artifact → Workout → Demonstration. The goal is incompressible content—experiments you couldn't find elsewhere because they emerge from the specific intersection of my plateaus, my intellectual frameworks, and this particular tool.

THE UNDERLYING PHILOSOPHY.

"You can't be wrong if you are vague enough."

Most guitar content is vague. "Learn your modes." "Practice scales over changes." "Feel the music." These aren't wrong, but they're not actionable. They leave the hard work of translation to the student.

String37 is an attempt at precision. Each experiment isolates a specific problem, proposes a concrete intervention, produces a tangible artifact, and measures a definable outcome. The intellectual scaffolding (philosophy, systems theory, finance analogies) isn't decoration—it's a forcing function for clarity. If I can't explain why this approach works using first principles, I probably don't understand it well enough to teach it.

The Chord Tone Atlas is the first artifact. It won't be the last.