Augment, Don't Automate: How to Stay Valuable While Using AI

No, you're not going crazy. This is a bizarre time to be a software engineer. Personally, I once took great joy in showcasing my craftsmanship in efficient pipelines, clean code, and project scaffolding; however, AI now does 99% of this for me.

So...who am I?** I'm an engineer.** I build things for people using the Internet. The struggle I'm talking about today is staying out of AI's shadow - that is, augmenting your skills with AI instead of allowing them to be swallowed by automation.

From danger to opportunity

In their September 2025 Economic Index report, Anthropic defined two types of AI tasks: automation (replacing human cognition) and augmentation (supporting human cognition). In their analysis of 1,000 Claude conversations globally, they highlighted a significant uptick in automation (77% of all Claude requests) relative to augmentation (only 12%) over the last 12 months.

The report defined a clear danger for workers leaning in to automation:

If more advanced models simply expand the set of automated tasks, then the risk increases that workers performing such tasks will be displaced.

We all need AI automation in 2025, but start from a headspace of reinvention, which is the only defensible moat. Don’t just blindly automate your job; reinvent it every time you open Cursor or Claude.

A few traps to avoid:

Local Maxima - Focusing on AI automation tasks using technology from 2+ years ago (instead of searching for higher ground, new workflows, new abstractions, new user value). Prompt suggestion: "Optimize this code for new possibility."
Commoditization - Focusing on the same generic AI automation as everyone else, thereby flattening your own differentiation. Prompt suggestion: “What part of this workflow is truly proprietary knowledge that could be its own product?"
**Skill Decay **- Allowing AI automation to supersede your sense of taste, judgment, and basic programming know-how. Prompt suggestion: “Don’t give me the final answer! Walk me through your reasoning and ask where mine might differ.”

Augmentation strategies

The biggest advice I have is to orient on outcomes, not tasks. For example: replace “automate step X” with “increase outcome Y." *Examples: reduce time-to-insight, increase experiment throughput, raise shipment reliability, surface novel customer jobs. *This serves as my "AI cognitive exoskeleton" that I use for all Cursor, Claude, or Codex sessions.

I typically employ this strategy in the context of the following 30-minute human-in-the-loop framework:

**Pre-brief **(5 min): Think of this as the start-up script for your AI OS:

My goal is x; success looks like y; unknowns: z.

Diverge (10 min): Ask the model for new ways of thinking.

Operating within software engineering best practices and avoiding any ideas that require overhead, give me 3 simple, actionable ways to change the way I'm approaching this work, not just speed it up (e.g. delete x step, compress feedback loops, invert y data flow, redefine handoffs). Skip anything that requires overhead; we are looking for a individual approach to building/maintaining systems. 

Return a detailed analysis of each of the 3 approaches in a APPROACHES.md file.

Other examples of divergence to play with:

Option sets, not answers: “Give me 3 distinct ways to restructure this pipeline. For each: upside, risk, 1-hour test.”
Abstraction hunts: “Given this codebase, propose 5 new boundaries or contracts that would remove classes of bugs.”
Customer reframing: “Generate 10 fresh ‘jobs-to-be-done’ for this feature from support logs; map to experiments.”
Deletion design: “If we had to ship the same value with 50% fewer steps, what would we cut or merge? Draft the new SOP.”
Adversarial review: “Red-team this release plan for compliance, data lineage, and user harm. List mitigations.”
Converge (10 min): Ask the model to reduce down wide scope into something narrow and focused.

Read the APPROACHES.md file containing 3 competing approaches. Define a rubric and grade each on that rubric. Take the winning idea and render me a full plan from discovery to launch to accomplish my goal using the above success criteria.

During the convergence phase, here are a few questions I rely on:

What outcome am I moving? (not: which step am I speeding)
What could I delete if the model helped differently?
What’s the new affordance unlocked (new product, new insight, new user power)?
How will I know in < 1 day if this direction is promising?
What would make this unsafe or unethical, and how do I bound it?
Critique (5 min): Ask the model to adversarially review your plan: risks, ethics, failure modes, blast radius. Define this in a rubric for max explainability/reproducibility.

You are a code nerd. For each system/code review, rate (0-10) with receipts:
**Bugs & Footguns**: Edge cases, race conditions, memory leaks
**Security Oof Moments**: Injection flaws, auth bypasses, secrets in git
**Failure Modes**: What breaks? How bad? (severity × likelihood)
**Blast Radius**: Does one bug nuke prod or just one function?
Output: Total jank score, yikes list, fixes ranked by pain level.

When to automate

Remember, automation isn't bad. My recommended approach:

step 1: augment w/ AI - discover problems, reinvent the workflow, "set the stage".
step 2: automation w/ AI - writing tests, docs, pipelines, make things pretty.

Hopefully this was helpful in some way! Please leave feedback and look forward to a second post that just focuses on using AI to augment ideation.