marketerswiki
HomeCommunity
Resources
Learn
marketerswiki

The open playbook for performance marketers who build with AI.

Resources

  • SimpleCRM
  • Ads OS
  • Ledgeros
  • Tag Manager Engine
  • Skills
  • Playbooks
  • Vibecoding

Learn

  • Harness Engineering
  • Glossary
  • How-To Guides
  • Compare Tools
  • Claude Code For...
  • AI Tools Directory

Company

  • About
  • Community
  • Privacy Policy
  • Terms of Service

© 2026 marketers wiki. All rights reserved.

Built withClaude Code
A new approach to working with AI

Stop Prompting.
Start Harnessing.

The shift from asking AI for help to building the systems that make AI reliable, verified, and compounding.

Enter
The Problem

The same model. Different harness.
52.8% vs 66.5%.

LangChain ran the exact same model on Terminal Bench 2.0 — twice. The only variable was the harness. Score jumped from 52.8% to 66.5%. Same model. +13.7 percentage points. The model was never the bottleneck.

Agentic Laziness
The agent games your signal. If "done" means "tests pass," it deletes the failing test.
Self-Preferential Bias
The agent grades its own work. It always gives itself straight A’s — even on broken output.
Goal Drift
Over long sessions, the agent forgets what it was supposed to do. By turn 50, it’s working from foggy memory.

Most people are running agents with no harness. That’s not an agent — it’s a model with a prayer.

The Evidence

The environment matters more than the model.

13.7%
Performance gain from better harness, same model
LangChain, Terminal Bench 2.0
−80%
Tools removed from agent — performance improved
Vercel
$9 → $200
Solo agent: broken app → Full harness: working software
Anthropic
859
PRs merged in 30 days, AI wrote every one
Peter Steinberger
The Shift

Four stages. Each one wraps the last.

2022
Prompting
Context given as natural language. You type a request. The model responds. One shot. Every session is a roll of the dice.
2025
Context Engineering
Systematic construction of what the model sees. A 41% mistake rate dropped to 3% just by adding a comprehensive context document.
2026
Harness Engineering
The full system around the model: instructions, state, verification, scope, lifecycle. The 5 subsystems that make output reliable.
Now
Loop Engineering
The outer system that keeps an agent running, spawns helpers, verifies, and decides the next move. The prompt becomes the smallest layer.

Each stage doesn’t replace the last — it wraps it. The prompt doesn’t disappear. It becomes infrastructure.

“

I don’t prompt Claude anymore. I have loops running that prompt Claude themselves and decide what to do next. My job is to write these loops.

Boris Cherny · Head of Claude Code, Anthropic

You shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.

Peter Steinberger · 859 PRs merged in 30 days, AI wrote every one
The Domains

Nine domains. One mental model.

The same harness principles apply everywhere. The nouns change; the verbs don’t. Each domain has its own verification signal, its own human gate, its own loop.

Software Engineering
Verify: Exit code 0
Gate: PR review
Ghostwriting & Content
Verify: Humanizer + voice match
Gate: Always (subjective)
Marketing & Ads
Verify: ACOS, ROAS, TACoS
Gate: Big bid changes
Research & Analysis
Verify: Validator + verifier
Gate: Pipeline is the gate
Knowledge Management
Verify: Wiki-lint
Gate: Lint is the gate
Loops for PMs
Verify: Known-example evals
Gate: PM defines correct
15 Concrete Loops
Verify: CI green, score thresholds
Gate: Varies
Performance Marketing
Verify: CPA, ROAS, frequency
Gate: Creative approval
Lead Nurturing
Verify: Meeting-booked rate
Gate: Compliance + consent
The Architecture

Five subsystems. One harness.

1
Instructions · AGENTS.md
Startup path, rules, definition of done
2
State · feature_list.json
What exists, what’s next, what passed
3
Verification · init.sh / tests
Objective signal the agent can’t game
4
Scope · Feature dependencies
One thing at a time — no overreach
5
Lifecycle · session-handoff.md
Next session starts working, not exploring

Verification (highlighted) is the sole defense against agentic laziness. Four of five subsystems fight goal drift. Together they neutralize all three failure modes.

The Universal Loop
Observe → Choose → Act → Verify → Record → Repeat

The same six steps in every domain. The nouns change; the verbs don’t. If no new feedback can change the next action, it’s a one-shot — not a loop.

A prompt is a request.
A harness is a work system.

The people getting leverage from AI aren’t writing better prompts — they’re designing better systems. Skills compound. State files don’t. The harness is where the leverage lives.

14 lessons. 9 field guides.
One mental model.

From the foundational equation (Agent = Model + Harness) through the complete transition plan (how to never prompt again). Theory, applications, and the concrete systems for engineering, marketing, content, research, and more.

Enter

Passcode protected · Persists across all pages