Tweak Idea 2.0: better, faster, stronger (and new commands)

What changed in v2

Last week I shipped Tweak Idea — a Claude Code skillset that evaluates startup ideas across 14 weighted dimensions and gives you a scored report with concrete next steps for developing the idea. The v1 article walks through the original pipeline in full: the research stage, the 14 independent scoring agents, the hypothesis system.

The original pipeline — still the foundation in v2:

Tweak Idea pipeline

Some of the changes in Tweak Idea 2.0.0:

deterministic Python scripts to do most of the work that used to run inside the LLM
new commands for browsing, comparing, and improving ideas
new amazing report layout
profiles for your potential co-founders

The commands

In v2 the one /tweak:evaluate command becomes a family. Each new one is a step in the real loop of working on an idea: run it, look at it later, compare versions, get suggestions, find new things to try.

/tweak:evaluate

Start here. Pass an idea, get a full 14-dimension report back.

/tweak:list

Shows every idea that was evaluated so far.

/tweak:show

Opens a specific report in your browser. No re-running, just re-reading.

/tweak:compare

Diffs two runs side by side. It shows exactly what moved after you've iterated on an idea.

/tweak:improve

Reads your latest report and hands back concrete tweaks for the weakest dimensions.

/tweak:browse-hn

Scans Hacker News and surfaces posts describing real technology shifts. Feed the good ones into /tweak:suggest-from-hn.

/tweak:suggest-from-hn

Reads an HN post and generates startup ideas built on the shift it describes.

/tweak:evaluate "an app that helps restaurants reduce food waste"

Your HTML report is ready. Want me to open it?

JSON in, JSON out

In v1, the LLM did everything. It wrote the whole report in prose, hundreds of lines, on every run. It generated the HTML output word by word. Every stage picked its own format.

In v2, the LLM reads JSON files and writes JSON files. The shape is fixed up front by a schema. Once a JSON file is written, nothing can change it. This workflow is great for Python: we can use scripts wherever the LLM is not necessary, for example to build different report formats.

Three things change:

Reliability. Claude Code struggles with agent invocations that have to write a 700-line file. A 30-minute run no longer crashes halfway through a free-form report.
Speed. A full run takes about 10 minutes now, down from 30. What the LLM does in 20 seconds, Python does in 20 milliseconds.
Bounded failures. When one JSON file comes back broken, you point the LLM at that one file and ask it to fix it. Nothing else re-runs.

That last point is the one that matters most for serious Claude Code skills. The model is fast and creative, but it is not a replacement for code. Treat each LLM call in your skill as a function. Let code drive the pipeline. When the model fails, the failure stays in one place, and so does the fix. That is the pattern that made v2 work.