Iskra (Polish for spark) is a Claude Code skill that forces non-obvious solutions through a controlled research process. It prevents AI from taking the easy path.
In July 2025, Przemysław Dębiak "Psyho" — a Polish competitive programmer and multiple TopCoder champion — won the AtCoder World Tour Finals in Tokyo. He beat eleven of the world's best programmers and an OpenAI model. The only human to outscore AI.
When asked what gave him the edge, he said:
"AI lacks what I call research taste — it always goes for the easiest solution, while a human can push through a longer, harder path to reach something better."
If you work with AI regularly, you know this feeling. The model gives you answers that are correct, fluent, complete-looking — and completely average. Something's missing.
It's not a bug. By design, the model optimizes for:
- Probable — the most statistically common answer
- Safe — no risk, no surprises
- Quickly acceptable — looks complete, so it feels good
- Familiar — matches patterns from training data
- Finished — closes the topic instead of exploring it
This is exactly what kills originality, strategic depth, and real creative advantage.
Writing "be creative" or "think outside the box" changes nothing. The model understands these phrases stylistically, not operationally — and still defaults to the first comfortable average.
You can't give AI genuine research taste. But you can force it to simulate the behaviors of someone who has it.
Iskra designs a process where the easy answer is forbidden or insufficient:
- First, identify and block the obvious solutions
- Then, generate variants across a spectrum from conservative to seemingly absurd
- Separate generation from evaluation — so a weird idea doesn't die too early
- Attack each idea before recommending it
- Defend the strangest variant before dismissing it
- Only then choose a direction using hard criteria
You won't make AI have a spark. But you can make it not stop working before the moment where a spark has a chance to appear.
✅ Good fit:
- Copywriting, naming, taglines, campaign concepts
- Product, content, or business strategy
- Technical architecture decisions
- Any task where "correct" isn't enough — you need better than standard
❌ Skip it for:
- Routine tasks (formatting, summarizing, translating)
- Tasks with a single clear correct answer
-
Find your Claude Code config folder:
- Windows:
C:\Users\YOUR_NAME\.claude\skills\ - Mac / Linux:
~/.claude/skills/
- Windows:
-
Create the
iskradirectory and placeSKILL.mdinside:.claude/ └── skills/ └── iskra/ └── SKILL.md -
Restart Claude Code.
-
Invoke the skill:
/iskra
Type /skills after restart — iskra should appear in the list.
After invoking /iskra, Claude asks which mode you want:
QUICK or DEEP?
- QUICK — fast "sabotage → build" check, ~5 minutes
- DEEP — full 8-step process for strategic work
Example:
/iskra
> QUICK
> Task: Come up with a name for an app that teaches languages through AI conversations
The model will:
- Give you the obvious first answer ("LinguaAI", "TalkBot")
- Explain why it's mediocre
- Propose 3 less obvious alternatives
- Attack each of them
- Pick the best one with reasoning
Total: ~5 minutes, concrete output.
Example:
/iskra
> DEEP
> Task: Go-to-market strategy for a SaaS tool targeting small marketing agencies
The model runs 8 steps:
- Problem diagnosis — is this really about GTM, or about validating product-market fit first?
- Block the obvious — cold outreach + content marketing → FORBIDDEN as the sole approach
- Generate variants — from waitlist landing page (conservative) to software house partnerships (breakthrough) to "pay what you want for 3 months" (seemingly absurd)
- Attack each variant — what could go wrong
- Defend the strangest one — why "pay what you want" might outperform freemium in this segment
- Selection across 8 criteria
- Final recommendation — one direction, no hedging
- Concrete output — no filler phrases
Honest answer: it won't give AI genuine research taste. The model has no intuition, no obsession, no feeling that "this weird approach might be brilliant."
But it will force a simulation of the behaviors someone with that taste performs:
- doesn't trust the first answer
- looks for hidden assumptions
- tests variants instead of declaring answers
- defends weird ideas instead of killing them
- compares strategic advantage, not just correctness
iskra/
├── README.md ← this file (English)
├── README.pl.md ← Polish version
├── SKILL.md ← skill file to install (English)
└── SKILL.pl.md ← skill file in Polish
If you find this useful, especially for non-obvious use cases — I'd love to hear about it. If you find a situation where it doesn't work as expected — that's even more valuable.