Deep research structured like engineering work. Six phases, each with a clear role for you and a clear role for AI.

The measure of learning is not how much you read. It is how much you produce. Input that never becomes output is just consumption.

AI is a tool in this process, not a replacement for your thinking. Never let AI read for you, decide for you, or write in your voice without your direction. The moment you outsource the thinking, the learning stops.

Before starting, use AskUserQuestion to confirm which mode fits the goal:

If the user is unsure, suggest Quick Reference. Do not proceed until mode is confirmed.

Gather only high-quality primary sources. For a technical domain this means: recent papers that introduced key ideas, official model and product blogs from the labs, posts from the people who built the thing, university course materials from the last two years, and canonical "build it from scratch" repositories.

Not summaries. Not explainers. Not "top 10 things you need to know." Sources.

For each source: download it, convert to Markdown, clean, and file it into a structured research directory organized by sub-topic.

undefined

Work through the materials yourself. For each piece:

At the end of this phase, cut roughly half of what you collected. If you cannot decide what to cut, the material was not strong enough.

For key concepts, try cutting across multiple angles: how did this idea evolve historically? What is the opposite claim? What does it look like in practice vs. in theory? What breaks if you remove it? This kind of multi-dimensional interrogation builds a real mental model, not just a definition.

For fields with a research lineage, trace the intellectual genealogy: what problem was the foundational paper or idea responding to? What did it critique or improve on? Follow that chain backward two or three steps, then look forward at what built on top of it. Understanding a field as a sequence of problems being passed forward is more durable than understanding it as a set of current facts.

Claude's role here: explain, translate, answer specific questions. Not summarize or evaluate on your behalf.

Once you have a working mental model of the domain, write an outline for the article.

The outline must answer: what do you want to say, and what does your reader need to know to follow it? These are not the same question. Know your reader's level and write for that level, not above it.

For each section in the outline: note the source materials it draws from. If a section has no sources, either it does not belong in the article or you need to find the source.

The outline is a contract with yourself. Do not start filling it in until it is solid.

Work through the outline section by section. This is the repetition phase: revisiting material you already processed, synthesizing it into continuous prose, filling gaps as you find them.

You will end up with a long, somewhat rambling draft. That is correct. Do not edit while writing. Get it all down first.

If a section is hard to write, it usually means the mental model is still weak in that area. Go back to Phase 2 for that sub-topic rather than forcing through. The outline may need to change, and that is fine.

Now hand the draft to Claude with a specific brief:

Work through the suggestions yourself. Accept, reject, or modify each one. Do not accept blindly. This phase often surfaces things you missed in Phase 2, which means more learning.

Before accepting Claude's edits, run on the refined draft to strip any AI patterns that crept in during the refinement process.

Read the entire article yourself, not with AI. Read it as your target reader would: linearly, without jumping back.

Mark everything that feels off: unclear sentences, abrupt transitions, sections that drag. Fix them. Read again. Two full passes is the minimum.

When it reads clean from start to finish: publish it.

One concern to set aside: "what if no one reads it." If the content has substance, readers will find it. That concern is not a reason to skip publishing. It is a reason to make sure the content has substance.