Six Hats: A Decision Skill That Completes My AI Context Stack

I keep adding small skills to my agent setup, and I keep noticing the same gap: AI is great at collecting and synthesizing, but not at taking sides. When I ask for “thoughts on X,” I usually get a balanced essay where every paragraph hedges the previous one. That’s not a decision — that’s a Wikipedia entry with feelings.

So I built six-hats-skill, a Claude/Cursor skill that runs an Edward de Bono–style six-hats debate over a topic, in three rounds, ending with a moderated recommendation.

The point isn’t the framework. The point is forcing the agent into roles that disagree on purpose, and then making one of those roles synthesize a real call at the end.

What it actually does

You hand it a decision and an output path. It runs three sequential rounds with these roles:

White Hat — facts, known data, gaps in knowledge
Red Hat — gut feel, intuition, emotional reaction (no need to justify)
Yellow Hat — upside, best-case, why this could work
Black Hat — risks, failure modes, sharpest objections
Green Hat — alternatives, reframes, lateral options
Blue Hat — moderator; doesn’t argue, just synthesizes the final call

Each round sees the previous round, so positions can shift, sharpen, or concede. At the end, Blue Hat writes a final recommendation, key agreements, unresolved tensions, and next steps to a markdown file — the format my Obsidian vault wants.

Two examples are in the repo:

The React one is a good demo of the value: the hats actually disagree about whether the framework’s maturity is leverage or a trap, and Blue Hat lands on a real position (“treat your stack as a liability to be managed, not an asset to be defended”) instead of “it depends.” That sentence wouldn’t have come out of a single-prompt synthesis.

Why this slots into a stack of skills

I’ve been leaning on three other skills heavily, and six-hats is the missing fourth corner.

Skill	What it gives me	When I reach for it
last30days	Outside-in signal — what Reddit, X, YouTube, HN are actually saying in the last 30 days	Early scoping. “Is this even a real thing right now?“
deep-research	Citation-backed, multi-source synthesis with a real methodology	When I need to trust a claim, not just feel it
grill-me	Relentless interview about my plan, branch by branch	After I have a draft plan and need it stress-tested
six-hats (new)	Structured internal debate from six opposing lenses, ending in a recommendation	When I need to decide, not just understand

Each one answers a different question:

last30days → What’s the world saying?
deep-research → What’s actually true and well-sourced?
grill-me → Where is my plan weak?
six-hats → What’s the right call, given everything above?

If I only use the first three, I end up with a pile of evidence and no decision. Six-hats turns the pile into a position.

The loop I’ve settled into

For anything bigger than a one-day spike, the flow now looks like this:

last30days <topic> — quick scan of what’s being discussed right now. This catches things that aren’t in my head yet (new tools, hot takes, dead repos, recent regressions).
deep-research <topic> — I run this when last30days surfaces a claim I want to lean on. I want citations, not vibes, before I commit to a direction.
six-hats — I feed the topic plus the research into a debate. White Hat now has real facts to cite. Black Hat has real failure modes from production reports. Green Hat has real alternatives instead of made-up ones.
grill-me — once Blue Hat gives me a recommendation, I let grill-me tear into my version of the plan, branch by branch, until I can’t dodge a question.
Decide and write it down — usually a short doc in the vault with the Blue Hat synthesis on top, the grill-me transcript below, and links to the research.

The order matters. Six-hats before research is a vibes session. Grill-me before six-hats grills you on a position you haven’t actually picked yet. Putting them in order turns four general-purpose skills into a pipeline.

What changed when I added six-hats

A few things I wasn’t expecting:

Black Hat catches what I won’t. I’m naturally a Yellow Hat about my own ideas. Having the agent role-play a real devil’s advocate, with the previous round’s optimism in front of it, surfaces objections I would have rationalized away.
Red Hat is weirdly useful. “What does this feel like?” sounds soft, but it’s where I notice exhaustion, hype-fatigue, or excitement that I should pay attention to. The career example in the repo is mostly carried by Red Hat naming the fatigue.
Blue Hat forces a position. This is the whole game. The constraint that Blue Hat must end with a final recommendation — not a balanced summary — is what makes the output usable. I disagree with it sometimes, and that’s fine; disagreeing with a clear position is much easier than disagreeing with “here are some considerations.”
It’s better with research loaded in. Run six-hats cold and the hats invent stats. Run it after last30days + deep-research and the White Hat round reads like a brief. Same skill, very different output.

Things I’d tell myself before building it

Sequential beats parallel. I tried running the hats in parallel first to save time. The debate quality collapsed — each hat needs to see the previous ones to actually engage. The skill enforces sequential order on purpose.
Don’t let Blue Hat debate. Early versions had Blue Hat chiming in during rounds. It poisoned the debate — the moderator can’t also be a player. Now Blue Hat only shows up at the end.
Write the output to disk, every time. Same lesson as my content pipeline post — if it’s not a file in the vault, it didn’t happen. The skill always writes debate-{timestamp}.md to the path you specify.
Composability is the point. Six-hats is small on purpose. It doesn’t do research, it doesn’t grill, it doesn’t manage projects. It does one thing — structured opposing debate ending in a recommendation — so it can plug into whatever pipeline I’m in this week.

Try it

If you already have a Claude Code, Cursor, or ~/.agents/skills setup, installation is one line:

git clone https://github.com/juanallo/six-hats-skill ~/.claude/skills/six-hats

(Or ~/.cursor/skills/six-hats, or ~/.agents/skills/six-hats — same shape.)

Then ask your agent something like:

Run a six hats debate on whether we should migrate this app from REST to GraphQL. Save it to ./decisions.

Repo and full SKILL.md: github.com/juanallo/six-hats-skill.

If you’re already using last30days, deep-research, or grill-me, drop six-hats in next to them and try the full loop on a real decision you’re sitting on. The interesting moment is when the four skills disagree with each other. That’s usually where the real call lives.