Field guide · AI best practices · July 2026

What is no longer human work

How to use AI fully: refuse labor, keep judgment, and demand a decision surface. Companion to agent-centered coding workflows.

For students and builders Judgment · orchestration · labor Related: coding practices

1. The real lesson

“This is not human work. Use the tools. Bring me the decision surface.”

Example: an agent suggests reinstalling a Chrome extension by hand. Following those steps keeps you as the labor. The better move is to reclassify the task: machine chore = labor, then have the agent (or OS / browser automation) remove, reinstall, and test.

Fully using AI is not “AI while you still click everything.” It is assigning labor out of the room and keeping only the signature: the decision, the verify step, and the accountability.

2. Three layers of work

Judgment

Values, risk, relationships, “should we?”, certification of truth. Stays human.

Orchestration

Specs, constraints, verify, escalate, load the right skill. Human + light AI.

Labor

Clicking, drafting, searching, formatting, scaffolding, reconciling, first-pass analysis. Agent.

SOTA workflow = keep judgment, systematize orchestration, automate labor.

Skill amnesia is a common failure: multi-step work without plan mode or a named skill often produces garbage. That is not always “the model is bad.” It is often that you never loaded the process, so the agent invents one.

3. Not-human-work catalog

Domains where agent labor is normal if you still own the decision surface. For each area, practice naming: what is labor, what you must verify, what you alone sign.

Machine chores (browser, OS, installs)

Reinstall extensions, clear profiles, restart services, fix MCP / tool setup
Bulk rename/move, install CLIs, fix PATH, screenshot and verify UI

Rule: If the step is procedure on a computer, the agent should drive it. You only approve risky actions.

Software labor (beyond “features”)

Boilerplate, refactors, real tests, PR text, release notes
“Read the logs and give three likely causes”
Dashboards and admin views from the questions you care about

Agent coding (plan → harness → tests → review) is a subset of this layer.

Document and research labor

Pull criteria from references, build comparison tables, draft memos and emails
Example: criteria pass on a facility → human call to confirm → signed memo on why it was rejected

Rule: You do not need to have opened every page yourself. You need a verification plan and a signature of responsibility.

Communication labor (not relationship)

First draft when you do not fully understand the domain
Summarize a thread → best guess + open questions
Confirm with a person when stakes are real, then send

Pattern: agent = first analyst. You = editor and relationship owner. Guess → confirm → send.

Design / CAD / visual labor

Edit drawings, layouts, slides, UI mockups by voice or agent loop
“Change this constraint, regenerate views, export”

Spec + constraints + visual check. Hands on the mouse for taste, not every dimension change.

Ops and dashboards

Status boards for “what is true right now”
Reconcile sources into one review surface

Humans review dashboards. Building the dashboard is labor.

Life systems and learning systems

Mission, vision, weekly priorities, lesson scaffolds, tracking routines
Planning and logging systems that free humans for presence and coaching

Boundary: Care, attachment, and presence stay human. Scaffolding, logging, planning, and summarizing can be agent labor. “AI runs the pipeline; humans coach” - not “AI replaces the relationship.”

Meta-labor (highest leverage)

Write the skill or checklist so next time process loads automatically
Turn a successful run into a reusable skill
“If I did not name a skill and this is multi-step, stop”

4. What is still human work

Still human	Why
Final accountability	You sign the memo, the PR, the email
Relationship and trust	The confirming call; tone with people who matter
Taste and values	What “good enough” means for your product, team, or family
Risk and ethics	PII, legal, safety, money, chain of command
Verification design	What evidence would change your mind
Saying “not human work”	The meta-skill of refusing labor

Rule of thumb: If the cost of being wrong is social, legal, medical, or permanent, a human confirms. The agent still does the draft and the dig.

5. The protocol (use on every task)

Name the outcome in one sentence.
Classify: labor / orchestration / judgment.
If labor: assign tools, skills, or automation. Stop doing the clicking yourself.
Demand a decision surface: table, options, risks, draft, or diff - not the ordeal.
Verify proportional to stakes: glance / tests / call / dual-source.
Certify only what you own.
Compound: if you will do this twice, make a skill or checklist.

6. Phrases to practice

“This is not human work.”
“Give me the decision surface, not the ordeal.”
“Guess → verify → certify.”
“If I did not load a skill for a multi-step job, I am the bug.”
“I keep the signature; the agent keeps the shovel.”

7. What to try this week

Label five real tasks as judgment, orchestration, or labor.
Pick one labor task you usually do by hand. Force an agent (or skill) to do it. Speak the phrase: “This is not human work.”
Write the decision surface you needed: options, risks, draft, or diff.
Verify at the right level of stakes (glance, test, second source, or a human call).
Save one skill or checklist so you do not forget the process next time.

Win condition: three documented cases where you refused labor and still owned the outcome.

8. Pitfalls

Automating judgment - fake confidence on high-stakes calls
Skipping verify because the draft “sounds right”
Letting the agent thrash for 40 minutes when a two-minute constraint would unlock it
Privacy and sensitivity: health, kids, operational data - scope tools carefully
Status games: visible AI use while still doing invisible labor (clicks, installs, reformatting)

9. How this connects to coding agents

Coding agents	This guide
Agent stack, skills, plan mode	Same ideas as a full work operating system
Local models / tools	Labor vs judgment (hardware optional)
How you ship code	How you refuse labor across domains
Best practices for agents	Best practices for delegating to agents

You are not collecting more tools. You are building a habit of default refusal of labor plus honest verification. Related: AI coding now and next.