An Agent Skill that changes your question instead of answering it

Ask any assistant an open-ended question — should I build this, should I take this job, how do I position this product — and you get advice that is reasonable and slightly generic. Narrow the target. Differentiate. Build an MVP. Talk to users. Test demand.

None of it is wrong. It just never moves where you are standing.

Insight Compiler is an Agent Skill I published that treats that generic answer as the starting line instead of the finish. It writes the ordinary answer down explicitly, then spends the rest of the run climbing off it.

The loop

The skill runs seven steps every time:

1. Intent discovery — separates the surface question from the real goal, and names the question you haven't asked
2. Ordinary-answer baseline — writes out what a normal AI would say, explicitly
3. Domain lenses — looks at the problem through distant fields: military strategy, urban design, game design, religion, investment
4. Discovery operators — deforms the problem: invert it, swap who acts, shift the scale, hunt for the hole nobody is looking at
5. Hint bank — crosses lenses with operators into a dense list of raw material
6. Strategy candidates — 2-3 directions, each traced back to the specific hints it came from
7. Honest evaluation + one validation fork — ranks them without manufacturing a tie, then commits in advance to what each test outcome should change

Step 2 is the load-bearing one. An unstated ordinary answer is a ceiling the model quietly settles into; a written-down one is a floor you can push off. Step 7 is what keeps the whole thing from being a brainstorm — you end with one cheap test and a decision already made for either result.

The internal scan is wide, but the output is curated. It shows the strongest findings rather than dumping every lens it tried.

What it looks like in practice

Here is a real run. The question I gave it was my own, and it is a small one, which makes it a good demo — I published this skill and nobody was visiting the repository, so I asked what to do about it.

A plain assistant answered the question as asked: add more examples to the README, record a GIF, post it to X and Reddit, get the content ready before announcing. Solid, actionable, and it accepts every premise in the question.

The skill opened by rejecting the question:

Essential question: Not "how do I get more repo visitors," but "where does the target person first encounter enough value to feel something move in their thinking?"

Why this is non-obvious: The premise "low visits = insufficient explanation" conflates two different bottlenecks — the discovery/distribution problem and the comprehension problem.

Then, near the end, the assumption I hadn't noticed I was making:

The assumption that "GitHub repo visitor count" is the correct success metric, and that "hard to understand" is the cause of low interest, is itself questionable. For an agent skill that runs inside Claude, the person deciding whether to adopt it may never need to visit the repo at all — if they encounter a compelling output first.

Side by side:

Ordinary answer	Insight Compiler
Improve the README	Decide who it is for, first
Add 3-5 demos	Produce one complete example that lands with one person
Announce on social media	Distribute the output itself as the content
Grow visitor count	Create the first repeat user

It also graded its own three candidates honestly — flagging its most interesting one as "unverified whether target readers are actually where you post," and admitting the boring README fix was the only one executable today. The verdict: do the boring one because it is nearly free, but it is necessary, not sufficient.

The full transcript is in the repo, unedited, both answers side by side. It is long — that is the honest shape of the output, not a teaser.

Installing it

An Agent Skill is just a directory with a SKILL.md in it. For Claude Code:

git clone https://github.com/snc2work/insight-compiler.git
mkdir -p ~/.claude/skills
cp -r insight-compiler/InsightCompiler ~/.claude/skills/

Enter fullscreen mode Exit fullscreen mode

Restart, then /skills to confirm it loaded. Any agent that supports Agent Skills works the same way — copy the directory, keeping SKILL.md and references/ together.

You never invoke it by name. It triggers on the shape of the question:

Let me consult you on something. I'm deciding whether to quit my job to
build my own product.

Enter fullscreen mode Exit fullscreen mode

Users are complaining about our paid plan. Should we lower the price?

Enter fullscreen mode Exit fullscreen mode

I have an idea for a product, but I'm not sure it's different enough from
what's already out there.

Enter fullscreen mode Exit fullscreen mode

There is a second version, InsightCompilerPersonal, that keeps a local profile at ~/.insight-compiler/profile.md. It records only working habits — your known constraints, whether you actually run the validations you agree to, the angle you always enter from — and never which ideas you liked, so it grounds the executability judgment without building you an echo chamber. Install one or the other, not both; they trigger on the same prompts.

When not to use it

A full session is thousands of words. That length is deliberate, and it makes this the wrong tool whenever you want a quick factual answer, some code, or a decision you can act on in the next five minutes.

It also cannot check reality. When it says a market gap exists, that is a hypothesis produced by reasoning, not research. Treat novelty claims as things to go verify.

And if you already know what you are doing and just need it done, an ordinary answer is the correct answer. This skill is for the questions where you suspect you are solving the wrong problem.

One note on the example

There is something circular about the demo above, and it is worth saying out loud: the run told me to stop polishing the repository and put a finished output where readers already are. This post is that instruction being carried out. Whether it worked is measurable, and the session pre-committed to what each outcome means — reaction means keep producing transcripts, silence means the problem is not presentation and I should go narrow with one real user instead.

Repository: https://github.com/snc2work/insight-compiler (MIT)

If you try it, I would like to know what you pointed it at.