Get ideas your AI would never have given you
Ask any AI for ideas and you get the same obvious handful, reworded. This is the technique I use to break that open. Here's a real one: I asked for ways to get more paying hosts onto DailyPB, my pickleball platform — same question, answered both ways:
- Email underfilled hosts to "boost" their session
- Nudge sold-out hosts to promote the next one
- Onboarding drip for new hosts
- Win-back emails to lapsed hosts
- Referral ask: "invite another organizer"
- Make new hosts out of cancellations — auto-broker a paid sub when one falls through
- Run a reverse auction on dead court time — venues post empty slots, players claim them as hosts
- Scrape organizers already running groups off-platform, recruit them on autopilot
- "Fill guarantee" that promotes a session itself and only bills the host if it works
Same model, same brief. Every "before" idea was "email a host to spend money." The "after" ideas change where hosts come from.
Why your AI keeps handing you the obvious stuff
It's not that the model is dumb — it's that it's built to give you its single most-likely answer. So when you ask for "8 ideas," it doesn't explore 8 different regions of the idea space. It gives you its one top answer, sampled 8 times, with the wording shuffled. They look different. They're the same idea wearing eight outfits.
And the usual escape hatches don't help: regenerating gives you more reskins, the "make it more creative" dial just varies the words, and "act as five experts" gives you five voices saying the same obvious thing. None of them get you out of the obvious zone — which is exactly where the boring ideas live.
Why the "after" ideas are actually different
This is the part that matters. It's easy to make a list look more creative. The real test is whether the ideas are structurally different or just the same idea in fancier clothes. On the DailyPB run, the obvious ideas all shared one shape: "detect something about an existing host, email them to spend money." Even the "be more original" attempt only added cleverer triggers — waitlist depth, weather, milestones — the same machine with a fancier sensor.
The tail left that cluster entirely:
- Manufacture hosts from failures, not successes — when a session cancels, auto-broker a paid sub from a pool of vetted players, and that sub becomes a new host. Everyone else mines the wins; this mines the cancellations.
- Flip supply and demand — a reverse auction where venues post dead court time and players claim it as a session they host. Court supply becomes host supply.
- Recruit demand-first — generate a "ghost session" with no host, let players sign up to prove demand, then tell the top signups "this fills itself, want to run it?"
- Go off-platform — a scraper that finds organizers already running groups on Meetup and Facebook and recruits them, automatically.
- Remove the buy decision — instead of nudging a host to choose to promote, auto-enroll underfilled sessions in a "fill guarantee" that only bills them if it actually adds signups.
You don't reach a single one of those by regenerating the obvious list. That's the difference between AI that confirms your thinking and AI that expands it.
That DailyPB run isn't a one-off demo — this is a standing skill I reach for whenever I'm stuck circling the same few ideas, in my own work. The point isn't that every tail idea is a winner; some have real landmines, and naming those is part of the job.
The point is that two or three of them were directions I genuinely wasn't going to reach by thinking harder or asking again. That's the kind of leverage I wire into the businesses I work with — not a clever prompt, a repeatable system they keep.
→ Tell me the problem you keep circling — and I'll run it through this for you.
Here's the actual skill — take it
This is the part most people hide. I won't. A "skill" is just a plain-English Markdown file — no programming. Below is the real one I run, in full, lightly cleaned of my own business specifics. Copy it, download it, drop it into your own setup. The whole thing works because it's written down once and reused; there's no secret sauce I'm holding back.
# Idea Divergence
Break a model out of typicality bias and produce
**structurally different** ideas, not one idea reworded N times.
## Why this exists (the one thing to remember)
Aligned LLMs converge on their single highest-probability
answer. Asking for "8 ideas" returns **one idea sampled 8
times** with the wording rearranged. Regenerating, raising
temperature, and "act as 5 experts" all fail the same way —
they vary words, not the solution space. The fix (a 2025
paper, "verbalized sampling"): **ask for a distribution,
force the model to score each candidate's probability, then
return only the low-probability tail.** Reported effect:
1.6–2.1x more diverse output.
## When to use vs not
- USE for: marketing angles, automation/product ideas,
names/taglines, content hooks, offer structures, outreach
angles — anything where the answer space is wide and you
keep landing on the obvious.
- DON'T USE for: a constrained problem with one right answer;
"what should I do next" routing; grading an artifact. This
skill sits UPSTREAM — generate divergent raw ideas here,
then feed the keepers into the loop or the work.
## The procedure
### 1. Load real context (mandatory — skipping this is the #1 failure mode)
Thin prompts produce thin diversity ("context starvation").
Before generating, gather:
- The brief, stated concretely (the goal, for whom, what
success looks like).
- What already exists / has been tried / rejected — this
becomes the exclusion constraint in step 3. *Without real
exclusions, the tail collapses back toward the head.*
If the brief is one sentence with no context, ask 2-3 sharp
questions first (what exists, what's failed, what "good"
looks like) — don't generate on a blank slate.
### 2. Check for overfit-topic gravity
If the topic is something the model has seen ten-million
times in training (generic marketing tactics, "productivity
hacks," "engagement"), the tail is *still inside the well-
known cluster* and p<0.01 won't save it. Two fixes, apply
at least one:
- Exclusion constraint: "no idea that would appear in
mainstream [topic] writing."
- Specific frame: narrow the brief ("automated host-
acquisition machines that run on cron," not "marketing
ideas") or pull from underrepresented subcultures /
adjacent industries.
### 3. Generate with verbalized sampling
Run the model with a prompt shaped like this (adapt the
numbers and the exclusion to the brief):
> Generate ~20 candidate ideas for: [brief, with full
> context from step 1].
> For each candidate, estimate the probability that a
> typical LLM would produce it as a top answer (0–1).
> Exclude anything that already exists or has been tried:
> [the exclusions from step 1]. Also exclude anything that
> would appear in mainstream [topic] writing.
> Then return ONLY the 5 lowest-probability candidates —
> the tail. For each: the idea, a one-line why-it's-
> different, and its estimated probability.
Tune knobs to the job: pool size (20 default; 30–40 for a
wide space), how many tail ideas to return (5 default), and
how aggressive the exclusion is.
### 4. (Recommended) Run the three-arm contrast on a new kind of brief
To *prove* it's working (not just trust it), generate three:
1. Naive: "give me 8 ideas for X."
2. Naive-but-harder: same, "be genuinely original."
3. Verbalized sampling (step 3).
If arm 3 isn't visibly more divergent than 1–2, the brief is
probably overfit (revisit step 2) or context-starved (step 1).
On a brief type you've already validated, skip to arm 3.
### 5. Hand off — tail ideas are raw material, not finished plans
The tail trades safety for risk: the non-obvious AND the
unworkable both live there. Surface them honestly, flag the
landmines, let a human filter. Don't pre-kill a tail idea
because it has an obvious objection — name the objection and
keep the idea. Good next step for keepers: pressure-test them
against synthetic buyers, or develop the shortlist.
## Gotchas (both have bitten real briefs)
- Overfit-topic gravity (step 2) — the single most common
reason the tail looks boring.
- Context starvation (step 1) — the single most common reason
the tail looks generic.
- Don't confuse stylistic with structural diversity — five
different *voices* on the same idea is not five ideas. The
win is different *mechanisms / solution-space regions*, not
different tone.
## Reference
- Technique: "verbalized sampling" (2025 paper).That's the whole file — every step, every gotcha, the exact prompt. The leverage isn't in keeping it secret; it's in the judgment to run it well (loading real context, spotting an overfit topic) — the part that takes having actually done it.
Stuck on the same five ideas for your business?
This is one of a stack of techniques I bring into the businesses I work with. Tell me the wide-open problem you keep circling — the positioning, the angle, the offer — and I'll run it through this and send you the tail. 30 minutes, no setup on your end, no catch.
Email me the problem