mertcikla/tld-cli/create-diagram

Reference: When you need exact command syntax, flags, or a full walkthrough example, read references/tld-docs.md.

How tld thinks about architecture

In tld, everything is an element — a service, a database, a person, a class, a method, a module. Any element can have a view: a navigable canvas that holds child elements. Navigation works by drilling into any element whose view contains something interesting. There are no standalone "diagrams" to create separately — views emerge from the element hierarchy.

Connectors express relationships between elements in the same view. The view is inferred automatically from the elements' shared parent.

Roles (service, database, external system, person, etc.) are expressed as tags — not as distinct types. All elements share a single kind.

This is not traditional architecture diagramming. It is a navigatable atlas of the entire system — like a hyperlinked wiki where you can zoom into anything. At the top you see services talking to each other. Drill into a service and you see its modules. Drill into a module and you see its classes. Drill into a class and you see its methods, parameters, what each method calls, its parent classes, its subclasses.

The goal at full detail: a reader can start at the root view and follow drill-downs until they understand any piece of the system without ever opening a file.

The goal: a view that earns its existence

A view is only useful if it tells you something you couldn't immediately see from the file structure. The failure mode to avoid is a shallow view — a list of boxes with no connectors, or connectors with no labels, or elements that exist in isolation without showing who talks to whom, who owns what data, or where failures propagate. Boxes are cheap. The connectors are the insight.

Before moving to the next step, ask: would a new engineer reading this view understand how data or control actually flows? If not, add more connectors.

Working surface: diagram.sh

All tld commands go into a single bash script, diagram.sh, in the workspace directory. The script is your notes and your execution log — comments explain what you found, commands record what you built.

Create the script once at the beginning of Step 3:

#!/bin/bash
# Architecture diagram build script
set -e

For every batch:

1. Append a labeled block of commands (with a # === comment header) to diagram.sh
2. Run only that new block
3. Do a batch checkpoint before continuing to the next batch

The script is the complete, replayable record of every diagram decision. Never run a tld command outside it.

Batch checkpoint

Run this check after every batch of tld add or tld connect commands:

• New elements: Do any just-created elements also appear in other views (i.e., have a second placement)? Add them now.
• New views: Do any just-created elements' views connect to existing views via drill-down? No separate link commands needed — drill-down happens automatically when a child is placed inside an element's view.
• Cross-batch connectors: Are there interactions between elements from this batch and elements from a previous batch that aren't wired yet?

Append any missing connectors to diagram.sh and run them before moving on.

View density: the 15-element rule

Keep each view focused. Aim for ~15 elements per view; never exceed 25. A crowded view isn't more informative — it becomes a wall of boxes where the relationships disappear. If you're placing more than 15–20 elements in a single view, that's a signal to use tags, cluster, or both.

Tags: planes and boundaries within a view

Before clustering, consider whether tags can give the view enough structure to remain readable at a higher element count. Tags let a reader filter the view to a single plane, role, or boundary — making ~20 elements navigable where a flat list would not be.

Tags are applied at element creation time with --tag (repeatable):

tld add "Auth Service" --ref auth-svc --parent backend \
  --tag boundary:internal --tag protocol:grpc

Good uses of tags:

• Layer/plane: layer:api, layer:data, layer:infra
• Boundary: boundary:public, boundary:private, boundary:internal, boundary:external
• Protocol: protocol:grpc, protocol:rest, protocol:event
• Role: role:service, role:database, role:worker, role:gateway

Constraints — tags are not free:

• Each element should carry 1–3 tags maximum. More than 3 signals that the element is doing too many things or the tag scheme is too granular.
• Total tag count across the view must not exceed total element count. If you have 10 elements, you can assign at most 10 tags total across all of them — meaning most elements should have 0–1 tags.
• Do not tag every element for the sake of completeness. Tag an element only when its tag conveys something a reader couldn't infer from the element name and connectors alone.

Tags complement clustering — use both. Tags handle role clarity within a view; clustering handles size. A well-tagged view of 18 elements is better than a cluster of 8 that hides the relationships. A cluster of 5 with clear tags on each element is better than a flat list of 18 tagged elements where the groupings are invisible.

When and how to cluster

After mapping your inventory, count the elements you're about to place. If the count is heading above 15–20 even after considering tags:

1. List all the elements — write them out so you can see the whole set at once.
2. Find the largest cohesive group — look for elements that share a role, layer, or domain: all the auth-related pieces, all the data-layer components, all the event processors, etc.
3. Label the cluster by its role — give it a name that reflects what those elements do at that level of abstraction (e.g., "Auth Services", "Data Layer", "Event Pipeline").
4. Create a container element for the cluster — tld add "Auth Services" --parent root. This element's view will hold the cluster.
5. Place the clustered elements inside it — tld add "Token Validator" --parent auth-services.
6. Replace the cluster on the parent view with the container element — the parent now has one element for the whole cluster; drill in to see the details.
7. Repeat until every view sits at or below 15 elements.

The goal is not to hide complexity — it's to present the right level of detail at each zoom level.

Task Progress

Copy this checklist and check off items as you complete them:

• [ ] Step 1: Explore the codebase
• [ ] Step 2: Produce a subsystem inventory
• [ ] Step 3: Create diagram.sh and root elements — batch checkpoint
• [ ] Step 4: Add child elements per subsystem — batch checkpoint after each view
• [ ] Step 4a: Connect elements within each view
• [ ] Step 4b: 15-element rule (tags + clustering)
• [ ] Step 5: Audit shared elements — trace every view they appear in
• [ ] Step 6: Drill into every major subsystem (3–5 levels) — batch checkpoint after each view
• [ ] Step 7: Run tld validate
• [ ] Step 8: Ask user to run tld plan and give feedback; iterate if needed

Step 1: Explore the codebase

Before exploring, ask the user two questions:

1. How many views do you expect? (rough target is fine — e.g. "around 20", "as many as needed", "keep it under 10")

2. How deeply nested do you want the drill-downs? (e.g. "just top-level", "2–3 levels", "go as deep as possible")

Use their answers to calibrate the rest of the work. If they don't know, suggest options based on the codebase size once you've done a quick directory scan.

Shallow ~5–10 views, 1–2 levels deep — Services and their direct dependencies. Entry points, major layers, external systems. Nothing low level.

Medium ~10–30 views, 2–3 levels deep — Modules and packages decomposed. Key classes identified and placed. Major data and logic flows wired. Inheritance shown where architecturally significant.

Detailed ~50–200+ views, 4–6 levels deep — Every significant class has its own linked view. Each class view shows: __init__ parameters and their types, all public methods with connectors showing which methods call which, all private methods, the full inheritance chain. Every significant function shows its parameters, return type, and external calls. A reader should be able to navigate from the root view down to understanding a specific method's behavior without opening a file.

Exploration strategy

Start broad, then narrow:

1. Get the lay of the land — directory tree, top-level file structure
2. Find entry points — main.go, app.tsx, index.ts, server.py, etc.
3. Identify major layers — API, services, repositories, workers, frontends
4. Follow imports/dependencies to map how components connect
5. Check config files for external dependencies (DBs, queues, auth providers)

The output of this step is not a list of files — it's a map of connections. For each component you identify, note: what does it call? what calls it? what data does it read or write?

Step 2: Produce a subsystem inventory

Before creating a single element, write out an explicit inventory of every major subsystem you found:

Subsystem inventory:
- [name]: [one-line description] | calls: [...] | called by: [...] | shared deps: [...]
- [name]: ...

Do not proceed to Step 3 until every top-level directory or package is accounted for. A diagram that covers 60% of the codebase is worse than no diagram, because it creates false confidence.

Step 3: Create diagram.sh and root elements

Create diagram.sh, then append and run the root elements as the first batch. Root elements appear on the top-level canvas:

# === Root elements ===
tld add "Domain & Business Logic" --ref domain --tag layer:domain
tld add "Data & Persistence" --ref data --tag layer:data
tld add "Interfaces & Integrations" --ref interfaces --tag layer:interfaces
tld add "Platform & Infrastructure" --ref deployment --tag layer:infra

Append and run the next level as a second batch:

# === Level 2: major subsystems ===
tld add "Backend" --ref backend --parent domain --tag layer:api
tld add "Frontend" --ref frontend --parent interfaces --tag boundary:public
tld add "Storage" --ref storage --parent data --tag role:database

Batch checkpoint: Are there connectors between root elements suggested by your inventory? Add them now:

tld connect domain data --label "reads/writes"
tld connect interfaces domain --label "calls"

Step 4: Add elements and wire them together

Work one view at a time. For each view (parent element), append its children as a batch, run it, then immediately append and run its connectors before moving to the next.

# === Backend children ===
tld add "REST API" --parent backend --technology "Go" --ref api --tag protocol:rest --tag boundary:public
tld add "Stripe API" --parent backend --technology "Stripe" --ref stripe --tag boundary:external
tld add "Job Worker" --parent backend --technology "Go" --ref worker --tag role:worker

Batch checkpoint after elements: Do any of these elements also belong in other views? Add those placements now.

Step 4a: Connect everything

Append and run connectors for the same view immediately after its elements. View is inferred automatically:

# === Backend connectors ===
tld connect api stripe --label "billing"
tld connect api db --label "reads/writes"
tld connect worker queue --label "consumes jobs"

Labels should describe what the interaction does — "validates JWT", "publishes events" — not just "calls".

An unconnected element is a bug — either it's missing connectors, or it doesn't belong in this view.

Batch checkpoint after connectors: Any elements from other views that interact with elements you just wired? Append cross-view connectors now.

Step 4b: 15-element rule