Teach AI Agents to Think From Multiple Angles with Six Thinking Hats Skill

It is Thursday afternoon. You paste the launch brief into Cursor and ask the obvious question:

Should we delay the v2 release by a week to fix performance, or ship on Friday as planned?

The reply looks helpful. It cites one metric, names a risk, throws in a morale worry, suggests a compromise, and ends with:

it depends on your risk tolerance

You read it twice. You still cannot tell whether the performance data drove the answer or whether the model just wanted to sound balanced.

That is not a bad model. It is a mixed hat problem: facts, feelings, risks, upsides, and new ideas arrived in the same paragraph, so none of them are easy to audit.

The problem with one “helpful” answer

Hard decisions need more than multiple angles. They need angles kept apart long enough to trust.

Edward de Bono’s Six Thinking Hats is a form of parallel thinking: everyone focuses on the same kind of thinking at a step, then switches together. White is only facts. Red is only gut reactions. Black is only risks. Yellow is only benefits. Green is only new options. The colors are labels—the point is separation.

Once perspectives mix, basic audit questions get hard to answer:

What do we actually know?
What are we assuming?
What are people worried about?
Which ideas are new?
Why did we pick this recommendation?

In a workshop, a facilitator enforces that structure. In chat, there is no facilitator. The model’s default is to be useful in one pass—which usually means folding every hat into one polished block. You get the shape of rigor without the separation that makes rigor checkable.

What the same question looks like with hats

Take the launch question again:

Should we delay the v2 release by a week to fix performance?

Instead of one answer trying to do everything, a Full run at Standard depth might open with Setup, then walk the hats in order (abbreviated below):

Setup

Focus question: Should we delay the v2 release by a week to fix performance?
Mode: Full
Depth: Standard
Hat order: Blue → White → Red → Black → Yellow → Green → Blue

White

[KNOWN] Performance regression increased load times by 40%
[UNKNOWN] Impact on churn
[ASSUMED] Slower pages affect conversion

Red

Stated: The team is anxious about missing Friday’s ship date again
The launch might feel safer emotionally once performance is back in range

Black

Risk: Shipping poor performance damages user trust — Mitigation: Fix highest-impact bottlenecks first

Yellow

Benefit: Better launch experience — Condition: Performance improvements are noticeable to users

Green

Delay only performance-heavy features
Release to a smaller beta group first
Stage the rollout

Blue (synthesis)

Main tension: shipping momentum vs user experience
Recommendation: staged rollout
Next step: validate impact with a limited user cohort

Now you can see where the recommendation came from. You can disagree with it, challenge assumptions, expand one hat, or rerun a single section. The answer is easier to inspect instead of merely sounding convincing.

A fix you can install once

The six-thinking-hats Agent Skill encodes this structure for Cursor, Claude Code, and any host that loads Agent Skills. Instead of improvising “be thorough,” the agent runs one hat per section, then closes with Blue synthesis that may only stitch together what earlier hats already said.

Same focus question every time. Same rules about what each hat may add. A rerun next week should look like the same process, not a new essay.

New to skills? Read Your AI Agent Shouldn't Start From Zero Every Session first.

Try it on a real question

Install:

npx skills add ysskrishna/ai-agent-skills --skill six-thinking-hats

Invoke with the focus question in the prompt:

/six-thinking-hats Should we delay the v2 release by a week to fix performance?

Plain language works too—“parallel thinking,” “de Bono,” “six thinking hats”—when your host maps intent to the skill.

What you should see on the first run

Before any colored hat, the skill prints a short Setup block:

Focus question (one sentence)
Mode (which hat sequence; default Full)
Depth (bullets per hat; default Standard)
Hat order (the sequence it will follow)

If your brief is thin, it may ask up to three clarifying questions, then continue and label gaps. Even when you skip the details, mode and depth should still appear so you know what you got.

Then the hats run in order, each in its own section—not two hats in one block.

Hat	Job in that pass
White	Known, assumed, or unknown—no interpretation
Red	Stated or carefully inferred feelings only
Black	Risks and mitigations
Yellow	Benefits and what must hold for them
Green	New options, no scoring yet
Blue	Frame at the start; synthesize at the end without new facts or ideas

Blue bookends the run: it sets the question, then names tensions, a recommendation, and a next step drawn only from prior hats.

Pick a mode for the kind of decision

Most launch-or-invest calls want Full (default): Blue → White → Red → Black → Yellow → Green → Blue.

Mode	When to reach for it
Creative	Ideation; skips Black so criticism does not kill ideas early
Risk	Failure prevention; White and Black only, then Blue
Decision	Go/no-go; facts, risks, upsides—no Red or Green
Custom	Your order, still closing with Blue

Examples you can paste:

Six Thinking Hats: should we open-source the internal CLI?

Six Thinking Hats, Creative mode: newsletter topic ideas.

Six Thinking Hats, Risk mode: migrate auth this weekend?

Six Thinking Hats, Decision mode: ship the beta on Friday or delay by a week?

Match depth to stakes

Depth	Bullets per hat	Synthesis
Quick	2	3 bullets (tension, recommendation, next step)
Standard	3	Full Blue synthesis
Deep Dive	4–5	Richer synthesis for leadership readouts

Quick is enough for a hallway call. Deep Dive when the write-up has to stand alone in Slack or a doc.

How to tell the run actually worked

Skim for structure, not eloquence. The SKILL.md checklist is the source of truth; these are the checks readers notice first.

Structure (before you judge the prose)

Setup appears up front with all four fields: focus question, mode, depth, and hat order. If you did not specify mode or depth, they should still be stated explicitly (defaults: Full, Standard).
Hats run in the declared order, each in its own section—never two hats blended into one block.
The run closes with Blue synthesis after the last colored hat.

Per-hat formats

Hat	What to look for
White	Every bullet tagged `[KNOWN]`, `[ASSUMED]`, or `[UNKNOWN]`—facts and gaps only, no interpretation
Red	Stated feelings passed through verbatim; inferred feelings use “might,” “could,” or “may,” not asserted as fact
Black	Every bullet has Risk and Mitigation
Yellow	Every bullet has Benefit and Condition (what must hold for the upside)
Green	One distinct option per bullet; count matches depth (2 / 3 / 4–5); no scoring or ranking yet
Blue (close)	Names tensions (e.g. Yellow upside vs Black risk), a recommendation, and a next step—all drawn from earlier hats only

When to ask for a rerun

Hats mixed in one section (e.g. White, Black, and Yellow together)—say so and ask for one hat per section.
Blue overstep—new fact, risk, idea, or recommendation that no earlier hat supported. Point at the checklist and request a structured rerun.
Wrong order or missing Setup—the declared sequence was not followed, or the run skipped the opening Setup block.

When this skill is the wrong tool

Skip it for plain factual questions, execution-only implementation with no structured perspective pass, or a single-angle hot take when you do not want a fixed hat-by-hat sequence.

What to do next

Install with npx skills add ysskrishna/ai-agent-skills --skill six-thinking-hats.
Run a real question on your desk in Full + Standard mode.
If hats mix, link the SKILL.md and ask for a structured rerun.