Every line of PHP on shinobis.com was written with Claude. Every post on this blog was generated in a conversation with an AI. Every tweet, every translation, every technical audit passed through a language model before being published.
But delegating is not the same as approving. When I ask Claude to write a PHP function, I do not receive the output and push it to production. I receive the output and review it. AI created a review step that did not exist before. Before AI, I wrote the code and deployed it. There was no intermediate draft. There was no version to question. Now there is an output that is not mine, that I can evaluate with distance, and that forces me to consciously decide whether it is correct before adopting it.
That is not less work. It is different work. I moved from producing to evaluating. And evaluating requires judgment that producing does not always demand.
What I always delegate
PHP code. Every function, every SQL query, every .htaccess modification. Claude writes the first draft. I read it, question it, test it, and deploy it if it passes my review. AI-generated code is a draft, not a deliverable. If it works but I do not understand why it works, it does not ship until I understand it.
Translations and adaptations. Posts are generated in three languages in the same session. Claude produces all three versions knowing each one is independent. I review Spanish directly, verify the English tone, and validate the technical concepts in Japanese. The output does not publish without that review.
Social media content drafts. Tweets, LinkedIn posts, comment replies. Claude generates the draft with the tone and constraints of the channel (280 characters, contractions on Twitter, no contractions on the blog). I verify the message is correct, the data is specific, and the tone is mine.
Cross-platform technical audits. I give the same data to 3 or 4 AIs and compare responses. JSON-LD, UX, endpoints, llms.txt. AI does the analysis. I decide what to implement, what to discard, and what needs more investigation.
Proposals and quotes. Claude generates the structure, the non-technical language, and the professional formatting. I verify the scopes are correct, the prices reflect real work, and the terms protect both parties.
What I never delegate
Architecture decisions. What stack to use. What to build first. How to organize the database. How systems connect to each other. AI can suggest. But the decision to use vanilla PHP instead of a framework, to implement a graph search instead of text search, to connect posts with manual relationships instead of automatic tags, those decisions are mine. Because they imply long-term commitments that AI cannot evaluate without knowing my full context.
What to build. AI does not decide what tool to create, what post to write, or what client to accept. Those are strategic decisions that require business vision, personal values, and priorities that do not encode in a prompt. AI is excellent at executing tasks. It is still far less reliable at deciding which tasks matter.
Client relationships. Every conversation with a client has context that AI does not have. Relationship history. Sensitivities. What the client says versus what the client needs. AI can generate an email. It cannot decide the right tone for a client frustrated by a delay that was not my fault.
Final content curation. The last step before publishing is mine. Every post goes through a final read where I verify the tone is correct, the data is precise, there are no contractions in the blog, there are no dashes, and every claim is something I would defend if questioned in a comment. AI generates the content. I decide if it has my voice.
The editorial philosophy. Which topics deserve a post. What angle differentiates my perspective from any other blog. Which experiments are worth running. Which AI critiques I implement and which I discard. The direction of the blog is mine. The execution I share.
Delegating is not approving
The most common confusion I see in people starting to use AI is treating the output as a finished product. The output goes directly from the machine to the recipient without passing through human judgment. That is not delegating. It is abdicating.
Delegating is creating a division of labor where AI contributes speed and structure, and the human contributes judgment and responsibility. The review step is the added value. Without it, I am publishing the AI's judgment, not mine.
The line moved
At the beginning I delegated less. The first months I wrote the draft myself and asked Claude to improve it. I was afraid the output would not be "mine" if it did not start with me.
Over time the line moved. Now Claude writes the complete first draft and I review it. The result is better. Not because Claude writes better than I do. Because reviewing someone else's draft with distance produces observations that reviewing your own draft does not. You see the assumptions. You see the redundancies. You see what is unnecessary. It is easier to edit something you did not write than to edit something you wrote yourself.
But the line has a limit that did not move. Strategic decisions remain mine. The final voice remains mine. The responsibility for what gets published remains mine. That does not get delegated. Ever.
The rule
If AI generates something and I publish it without reviewing it, the responsibility for the error is mine, not the AI's. That is not an abstract principle. It is an operational rule that determines how I work every day.
Delegating is not trusting blindly. It is creating an intermediate step between the idea and the result where human judgment decides whether the output deserves to go out into the world. That step is the work. Everything else is the tool.