Claude Fable 5: Give It Goals, Not Tasks

There are two ways to run an AI.

Task mode: you give it a thing to do, watch it do it, and give it the next thing. Your job becomes sequencing work. Every output requires a human decision before the next input fires. The model is fast. You are the bottleneck.

Goal mode: you give it an outcome, define how success gets verified, and come back when it's done. The model works until the goal is met, checks its own output, reports what it did and what diverged, and surfaces only the exceptions that actually need you.

Most teams are running task mode with a fast model and calling it AI adoption. Fable 5, the latest release of Claude Code from Anthropic, is designed for goal mode. The gap between those two approaches is the gap between Level 2 and Level 3.

Why Task Mode Caps Out

Task mode feels productive at first. You write a prompt, the model returns something useful, you review it, you write the next prompt. The throughput is higher than doing the work yourself. Teams measure this in time saved per task and declare the pilot a success.

Then the ceiling arrives. You now have ten tasks running in parallel. Fifteen. Twenty. Every one of them produces output that needs a human decision before the next step fires. The people who were supposed to be freed up by AI are now spending their days managing AI output. You have hired a faster worker and promoted yourself to full-time supervisor. More AI equals more supervision overhead. That is the Level 2 ceiling, and it is structural.

The problem is not the model. The problem is the workflow design. Task prompts require constant human throughput to chain. As long as your unit of work is a task, your unit of bottleneck is a human.

What Goal Mode Looks Like in Practice

Fable 5 shipped two new Claude Code features that make goal mode concrete: /goal and verification workflows.

/goal keeps Claude working until a defined outcome is reached. You are not handing off a task. You are handing off an objective. Claude does not stop when it finishes one step. It continues until the goal is complete or until it surfaces a genuine exception.

Verification workflows define how success gets checked. You set the goal and you define what verification looks like. Claude does the work, then runs the verification against it, and returns a structured report: what was implemented, what matched the goal, what diverged and why.

The Anthropic Claude Code team described the intended workflow directly: after writing a spec, set a goal to implement the spec fully, then use a workflow to verify each part of the plan and prepare a report on what was implemented and whether anything differed. The model runs. You read the report. You make decisions on exceptions only.

The Verification Step Is the Unlock

Goal mode was theoretically possible before Fable 5. The practice fell apart at verification. Earlier models would complete a goal, report completion, and be wrong in ways that required a human to catch. You could not trust the self-reported outcome. You ended up reviewing every output anyway, which collapses goal mode back into supervised task mode with extra steps.

What changed with Fable 5 is that the model can reliably test its own work. It is not just completing the goal. It is confirming the goal was met against the criteria you defined. When something diverges, it surfaces the specific divergence with an explanation, not a blanket "done." The exception report is actionable. The standard path runs without you.

"The context window is your program. The model is the interpreter." — Andrej Karpathy, AI Ascent 2026

Karpathy's framing of Software 3.0 lands differently now. If the context window is the program, then your goal statement is the specification. Your verification workflow is the test suite. The model runs both. What you get back is a test report, not a stream of output requiring continuous approval.

What This Means for Your Team's AI Setup

Look at your current Claude deployments. Are they running on task prompts or goal prompts?

Task prompts look like: "Summarize this document." "Write a first draft of this proposal." "Review these three files and flag issues." Each one is discrete. Each one requires a human to receive the output and decide what happens next.

Goal prompts look like: "Implement this spec fully. Verify each section against the plan. Report what was completed and flag anything that differed." One instruction. The model works until done. You read the exception report.

The gap between those two is not a model problem. It is a workflow design problem. The model capable of running reliably in goal mode has existed since Fable 5 shipped. Whether your team is using it that way is a function of how your prompts were written, not what the model can do.

How to Redesign from Task to Goal

The redesign is not complicated. It is a discipline.

That is the compound engineering principle applied to AI direction. Every workflow you move from task prompts to goal plus verification pairs frees up a chunk of human attention. That attention can go toward the next redesign, or toward the work that actually requires human judgment.

The compounding is real. A team that has redesigned five workflows in goal mode has not just saved time on five tasks. They have restructured how AI attention and human attention interact across five entire functions. The sixth redesign costs them less than the first. The seventh, less still.

The Gap Is a Workflow Design Problem

The teams that will pull ahead in the next twelve months are not the ones with access to better models. Every company has access to Fable 5. The advantage goes to the teams that redesign their highest-value workflows for goal mode before their competitors do.

That redesign is not a technical project. It is an operator decision. Someone has to look at each repeating workflow, understand what the actual goal is, define what verification looks like, and rewrite the prompt architecture from task sequences to goal plus verification pairs. That work does not happen automatically. It requires someone who understands both the workflow and the model capability well enough to close the gap between them.

Your AI can run until done. Let it.