Context-Window Task Decomposition
Break jobs into bounded tasks that an AI can hold, judge, and complete.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 98%
Context-Window Task Decomposition treats a job as a portfolio of smaller activities rather than asking whether AI can replace the entire role. First, observe how the worker spends time and identify recurring tasks such as screening candidates, comparing documents, or drafting outreach. Then estimate the information each task requires and determine whether its inputs and expected output can fit inside the model's working context. Tasks that fit are structured with explicit instructions and delegated individually, while humans retain coordination, relationship management, exception handling, and context that spans the full role. The key measure is the proportion of working hours that can be delegated, because automating enough constituent tasks can substantially alter staffing even when the AI never performs the whole job autonomously.
Origin
Nathan Labenz developed this mental model while discussing how GPT-4 changes recruiting and other knowledge-work roles on Marketing Against The Grain.
Core principles
- 01A whole job usually contains more context than a model can handle.
- 02Many constituent tasks fit within the model's working memory.
- 03Delegation depends on structuring the input, not replacing a person wholesale.
- 04Evaluate work at the task and time-allocation level.
How to run it
- 1
Observe the Role
Create a time log showing what the worker actually does rather than relying only on the job description.
Pro tip Capture both frequent tasks and time-consuming irregular tasks.
- 2
Atomize the Work
Split the role into outcomes that can be completed and evaluated independently.
Pro tip Use verbs such as compare, classify, summarize, draft, or recommend.
Watch out Do not label an entire multi-stage responsibility as one task.
- 3
Calculate Context Needs
Identify the documents, instructions, history, and output space required for each task.
Pro tip Include the model's response in the context allowance.
Watch out A capable model still cannot complete a task whose necessary context does not fit.
- 4
Structure Delegable Tasks
Package qualifying tasks with clear inputs, decision criteria, and output formats.
Pro tip Begin with routine, documented activities that have obvious success criteria.
- 5
Redesign the Role
Assign the remaining coordination, judgment, and exception work to humans, then reassess staffing and throughput.
Pro tip Measure hours changed rather than counting automated task names.
Watch out Automation can reduce future hiring needs even when no current employee is directly replaced.
In the wild
A recruiting team separates candidate-role matching, personalized outreach, interviews, stakeholder coordination, and final judgment. GPT-4 receives a resume and job description to assess fit, then drafts outreach for candidates who pass the criteria. Recruiters focus on conversations, exceptions, and hiring-manager alignment.
→ A smaller team processes more candidates without pretending the model can perform the entire recruiter role.
Common mistakes
Testing the Whole Job
Asking whether AI can be a recruiter or doctor hides the narrower tasks it can already perform.
Ignoring Input Size
A task may be conceptually suitable but impossible when its required documents exceed the available context window.
Counting Tasks Instead of Time
Automating several minor tasks may matter less than automating one activity that consumes most of the workday.
Is it for you?
Best for
Leaders redesigning recurring knowledge-work roles around AI assistance or automation.
Not ideal for
Highly fluid roles whose tasks depend continuously on undocumented organizational context and original judgment.
From the transcript
“I think is the answer is kind of no, because jobs require more context than the AIs can handle.”
“However, what I think people do miss then when they say no to that first question is how many of the tasks that constitute jobs…”
“if you were to really take a time log of what is this person doing as they do their job of a recruiter, and you…”
From the episode
GPT-4 Beta User Reveals What Jobs It Will Destroy In 2023 with Nathan Labenz (#103)
Nathan Labenz