When you ask an AI a question, you get a first-pass best guess. Optimized for seeming complete. Optimized for the expected question, not your specific one.
The synthesis engine is a different approach: instead of one question and one answer, you run a sequence of passes. Each one builds on the last. The final output — the crystal — bears almost no resemblance to the first-pass answer.
The difference between one answer and a synthesis: roughly the difference between meeting someone for the first time and having worked with them for six months.
The Five Phases
Phase 1 — Braindump
Get everything out without filtering. Yes, include your pre-existing conclusions, your confirmation bias, your gut hypothesis. You're fully inhabiting the question before compressing it. The braindump should feel like it's going too far in one direction — that's right.
Give the AI: the question, all relevant context, your gut hypothesis, what a great answer would let you do.
Phase 2 — Extract
Compress the braindump. What does this actually mean? What's the real question underneath the stated question?
This is the inflection point. The extraction often reveals that the question you started with is not the question you were asking. Present the extraction to yourself: does this match what you actually meant? If not, revise and go back.
Phase 3 — Accumulate
Read everything relevant. This phase is read-only — you're gathering, not drawing conclusions. Add context, examples, competing views, data you've been avoiding. Maximum relevant material before synthesis begins.
Phase 4 — Steelman
Go back through everything assuming it's wrong. Write the strongest countercases. Where does the thesis break? What would a rigorous critic destroy first? What has the shortest half-life?
The psychoflexibility requirement: the steelman only works if you're genuinely willing to change direction based on what it finds. If you're doing it while already committed to the conclusion, you're performing it. The test: did the steelman change anything?
Phase 5 — Crystal
The final synthesis. The sharpest, most compressed version of everything developed across all passes. Not the conclusion — the current best synthesis, held with appropriate confidence. New evidence can update it.
Directing Passes: The "What's Alive?" Heuristic
After each pass, before starting the next: What was most alive in this pass?
Not "what was most important." Alive. What had an edge? What felt like it was pointing at something unexplored?
That element — not the outline, not the original question — should direct the next pass. Plans produce competent coverage. Following what's alive produces discovery.
Knowing When to Stop
The synthesis halts when the last two passes are both worse than the pass before them. Not when two passes are similar — similar passes sometimes reopen. When two consecutive passes have produced less new structure than the pass before them, the peak was already reached.
Practical signal: When you read your latest pass and feel like you're mostly restating what you said two passes ago — even with useful detail added — the peak was two passes back.
The output is the peak pass, not the final pass.
A Worked Example
Starting question: "I want to write about why most people use AI badly, but I can't figure out what I'm saying."
Pass 1 braindump: "The problem is everyone has access to the same AI and gets mediocre results... there's something about treating AI as a tool vs a collaborator... what I'm really annoyed about is that people ask AI for answers when they should be using it to think..."
Most alive line: "people ask AI for answers when they should be using it to think."
Pass 2 — Extract, follow the alive line: "Central claim: the problem is frame, not capability. People are asking AI to give answers when the better use is thinking through the question. The question formation is the work."
Most alive: the distinction between outsourcing thinking vs using AI to improve your own thinking.
Pass 3 — Accumulate + develop the distinction: "Outsourcing: ask 'should I take this job?' AI says yes/no. Using AI to think better: ask 'what am I missing about this decision?' You do the synthesis; AI provides material. Who holds the model?"
Pass 4 — Steelman: "Counter: sometimes you just need an answer. Stronger: this post says HOW to diagnose the problem but not how to fix it. Without a technique, this is just criticism."
The steelman found the actual gap — the post needs a concrete technique.
Pass 5 — Crystal: "Post argues: most people use AI as an answer engine. The key move is not 'what's the answer?' but 'what's the next question?' Here's one specific technique for doing this. Here's what a conversation looks like when you run it."
The crystal is completely different from the braindump. That gap is the value.
When to Use It
- Questions you've been stuck on for more than a few days
- Strategic decisions with irreversible consequences
- Writing projects you want to be genuinely good (not just done)
Not for: quick lookups, single-pass tasks, tight deadlines where 5+ passes aren't possible.
The tell: if a first-pass answer feels flat and you know more about the question than the answer contains, the synthesis engine is the right approach.
Practice
Run a 5-pass synthesis on one question you're currently stuck on. Write a companion analysis note after each pass (what appeared, what's alive, what direction to take next). Produce a crystal.
Completion criteria: A synthesis folder with 5+ passes, analysis notes, and a crystal. The crystal is materially different from the braindump.