AI Thinking Patterns Compared

TL;DR

Three distinct AI thinking patterns showed up in the same ninety minutes. Rod opened with the go-for-the-jugular pattern, asking the model to leap from a strategy matrix to ten LinkedIn posts in a single move. Catherine has been running a mycelial accumulation pattern, feeding context into long-lived sessions until a taxonomy forms underneath. Kent taught and demonstrated the sculptor pattern, which builds a layered Content Engine where every layer becomes context for the next.

Only one of these scales. The other two are useful, but they each have a failure mode you can predict in advance. The session is a clean lab specimen of why mental models matter more than tooling.

The three patterns, side by side

The patterns are not personalities. They are positions on a maturity curve, and most founder-operators are stuck somewhere on it without knowing.

Pattern 1: Go for the jugular (Rod's first attempt)

Rod started with a Massimo Matrix, a beautiful grid of broker issues mapped to stakeholders. Real strategy work, real proprietary IP. He pasted it and asked Claude for ten LinkedIn posts.

The output was usable. Kent's word for it: "totally usable, but it's not exactly what you're asking for."

That gap between usable and what you actually want is the entire lesson. The model can always produce something. It cannot produce the right something without enough context built up between the source material and the asked-for artifact. When you skip the layers, you get content that reads like a smart intern's first draft, not like Rod.

The diagnosis is simple. Rod's matrix is dense and rich, but it is not yet thinking. It is structure. The model has to do all the thinking in one pass: interpret the matrix, pick an angle, choose a voice, build a hook, write a post. Five jobs in one prompt. Fidelity bleeds out at every step.

The fix Kent prescribed:

1. New project, named for the purpose.
2. Drop the matrix in as a file, not as inline context.
3. Switch to a stronger model for the heavy lift. Sonnet for chat, Opus for ground truth generation.
4. Ask for one 250-word essay per cell, using writing-copy and the bespoke "ride like Rod" voice skill.
5. Treat those essays as your anchor content. Never ship them.
6. Open a new chat in the same project and ask for one LinkedIn post per essay using writing-linkedin-post.
7. Open another new chat and ask for X variants using writing-x-article.

That is sixteen prompts where Rod tried to use one. It looks like more work. It is not. The token cost across all sixteen prompts is lower than the single prompt would be after three rounds of "no, try again."

Pattern 2: Mycelial accumulation (Catherine's emerging method)

Catherine's pattern is harder to see because it works most of the time. She has been feeding context into Claude across many sessions, asking it to remember, asking it to build files, letting a taxonomy of capital intelligence problems grow underneath the conversations like mushrooms above a mycelium network.

She named the pattern herself, by accident, when she described "this body under the ground" with "fruiting bodies coming up." She was talking about her clients' business problems. She was also describing her workflow.

The strength: she now has organic context that no one could have specified in advance. The model has watched her thinking evolve. The taxonomy is emergent, not designed.

The risk: at some point the context window or the project memory will degrade, and she will not see it happen. She will get a confident answer that drifts from her real position. Kent caught a glimpse of this when she said "I haven't run out of credit, but I was doing that and getting compacted."

Compaction is the symptom. The cure is the same as Rod's, but applied to a long-running thinker rather than a content producer. Kent's prescription:

1. /clear the current session.
2. Open a new project.
3. Point Claude at active-projects/, completed-projects/, and org-config/.
4. Ask it to read everything she has done and synthesize.
5. Ask for a scoring rubric to prioritize which mushrooms to surface to clients first.
6. Optionally pull analyzing-text for a unity across the whole corpus, the one-or-two-sentence distillation that compresses 38 gallons of thinking into a single drop.

That last move, the unity, is the one most founders will never reach without help. It is syntopical reading made tractable by the model. You point at adversarial pairs (devil's advocate and steel man, two different stakeholder views, two different industries) and pull the common root. That root is almost always more valuable than either input.

Catherine's mycelium pattern, properly cleaned up, becomes a research asset. Left alone, it becomes context debt.

Pattern 3: The sculptor (Kent's Content Engine)

Kent's pattern has three commitments that the other two do not.

Layers, not leaps. The Content Engine moves from idea to anchor to distribution to promotion. Each layer is built once and reused many times. The matrix becomes essays. Essays become blog posts and LinkedIn and X. Top performers earn YouTube and ad spend. The cell is the unit of work, not the campaign.

Project memory is real and isolated. Kent kept correcting Rod into new projects and new chats, on purpose. Project memory is the single most underused feature in Claude. It persists what matters. It excludes what doesn't. The founder who learns this saves 60 to 80 percent of their tokens without losing a single capability.

Skill chains compound. The session demonstrated writing-copy × writing-linkedin-post × using-octalysis producing eight behaviorally-distinct hook variants from a single LinkedIn post that already passed quality. Eight by however many essays equals thousands of testable variants from one matrix. That is not content production. That is a behavioral test bed.

The sculptor metaphor is doing real work here. A bricklayer assembles small pieces into a large piece. A sculptor reduces a large block until the shape inside is revealed. Kent's pattern is reductive: build a dense block of context, then chip out the artifact you actually want.

This is the pattern that scales. It scales to teams (Catherine's question about extending to her org). It scales to onboarding (the "agentize yourself" idea for the new SDR). It scales across channels because the essays are the source, not the LinkedIn post or the X post.

What the session actually proved

Three things, none of them obvious until you watch them collide.

First, the same model produces wildly different output quality based on workflow shape, not prompt cleverness. Rod's first attempt and Rod's second attempt used the same Claude. The difference was the layered context built between them.

Second, project memory is the unit of leverage, not the prompt. Most "Claude is broken today" complaints are context-window pollution. The fix is structural, not verbal. New project, new chat, same project for chains that share context.

Third, skill chains beat skill stacking. Loading three skills in one prompt produces a blur. Loading them across a sequence (essay first, then LinkedIn, then octalysis hooks) produces eight crisp variants. Sequence matters. Order is a load-bearing detail.

Adoption recommendations

For Rod:

• Run the full Content Engine once on the Massimo Matrix. One project, three chats, three artifact types. Track output quality vs. effort.
• Fold "ride like Rod" into writing-copy rather than keeping it separate, so it inherits the writing-copy directives automatically.
• Use using-octalysis on top performers, not on every post. Eight hooks per winning post, not per post.

For Catherine:

• Run the look-back ritual Kent described before her trip to Portugal. Fresh project, point at all prior work, ask for synthesis plus scoring rubric.
• Pull a unity across her capital intelligence taxonomy. The drop is the headline of the document she is building.
• Add applying-systems-thinking for the mycelium framing. Meadows and Goldratt together turn the "fear on both sides of a transaction" insight into a leverage point map.

For Kent:

• The "sculptor not bricklayer" line is landing repeatedly. Make the YouTube video. The metaphor does heavy teaching work in one sentence.
• The Content Engine is implicit knowledge across many sessions. Codify it as a reference under designing-ai-workflows or as its own skill. It is the single most repeated teaching pattern this month.
• The "agentize yourself" pattern Catherine surfaced (the SDR talking to a Catherine-as-agent before bothering the human) is a packageable Tier-2 outcome. Worth a dedicated skill.

Closing observation

The three patterns map to a developmental sequence, not a taxonomy of users. Most founder-operators move through them in order. Go-for-the-jugular when they first try AI. Mycelial accumulation as they get comfortable. Sculptor when they finally treat the model as a capable colleague who needs a workspace, not a vending machine.

Next step: If Rod runs the full Content Engine pass on the Massimo Matrix this weekend, the output will validate or invalidate the layered approach against his existing voice. Either way it produces a usable training set for refining the "ride like Rod" skill.