Skip to content

Prompt complexity is a 7-rung ladder with a hard line between Level 3 (single-agent) and Level 4 (multi-agent) — pick the lowest rung that does the job reliably, not the cheapest that almost works.

Guide

How to Calibrate Prompt Complexity

The ladder

Seven rungs. One hard line in the middle.

L1

High-level intent

Single instruction, no structure. "Fix the typo in the README."

single-agent

L2

Workflow

Multi-step, linear. Check → identify → fix → commit.

single-agent

L3

Control flow

Branching, conditionals, loops. for each file in docs/ …

single-agent

L4

Delegation

Orchestrator routes to sub-agents. Multi-agent begins here.

multi-agent

L5

Higher-order

Prompts generate prompts — Phase Gate with Artifact Extraction.

multi-agent

L6

Template meta-prompt

Runtime {{VAR}} injection across many agents.

multi-agent

L7

Self-improving

Agent rewrites its own prompt from observed performance. Rare.

multi-agent

Each rung is a superset of the one below — and a multiple of the cost. The line between L3 and L4 is where design requirements change qualitatively, not just quantitatively.

The L3 → L4 inflection

Crossing to multi-agent is a one-way door.

Levels 1–3

One agent. One output. One invocation.

  • A single Command carries the whole task
  • Context and attention live in one place
  • Debuggable by reading one transcript
  • No handoff discipline required

Cheap to author. Ceiling: what one agent can hold in context and attention.

Levels 4–7

Orchestrator + sub-agents. New machinery.

  • Agent Delegation Hierarchy to route work
  • Per-agent skill composition
  • Subprocess handoff discipline
  • Domain Locking to keep multi-agent safe

Expensive to author. Only worth it when a single agent can't carry the load.

Don't cross the line because L4 feels sophisticated — cross it because L3 demonstrably fails.

How to pick a level

Lowest rung that does the job reliably — not at all.

Simple one-shot task?no structure, no branching

Level 1–2

·

Inline prompt in a Command

Recurring task with known steps?same shape every time

Level 3–4

·

Command template

Multi-phase SDLC work?plan → build → test → review

Level 5

·

Phase Gate with Artifact Extraction, chained in an ADW

Cross-team agentic system?many agents, shared structure

Level 6

·

Delegation Hierarchy + Metaprompt Variable Injection

Evolving long-lived agent?gets better over time

Level 7 (approx)

·

Level 6 + Mental Model updates

The rule is reliability, not capability. If L2 can do it but fails one time in ten, you're at L3.

Both directions cost real money

Under-engineering fails silently. Over-engineering burns tokens and hides bugs.

Too low for the taskunreliable output

L1 used for →

"Refactor the auth module and update all callers."

Misses callers, mangles edge cases, no phasing. Ships broken because the prompt had no room to branch or verify.

Too high for the tasktokens burned, bugs hidden

L5 used for →

"Fix the typo in README."

Three agents, an artifact handoff, a meta-prompt. 20× the cost, and a bug in the orchestrator now hides the trivial edit.

Most ADWs sit at L4–5. The Code Factory's /planning-products pipeline is L5 — a prompt that generates other prompts.

Where to go next

Three neighbor guides — one per altitude.

Any level

how-to-prompt

Writing a single-level prompt well — structure, specificity, failure modes at one rung.

L1–L7

Mid ladder

how-to-metaprompt

The mechanics of L5–L6 — artifact extraction, phase gates, variable injection.

L5–L6

Upper ladder

how-to-build-metaagents

Multi-agent architecture for L4 and beyond — delegation, skill composition, domain locking.

L4+

This guide tells you which rung. Those three tell you how to build well at that rung.

The discipline

Pick the lowest rung that does the job reliably — not the lowest that almost works, not the highest that feels impressive.

Under-engineer and you ship unreliability. Over-engineer and you burn tokens while hiding bugs behind orchestration. Calibration is the entire craft — the ladder is just the instrument.