An independent second opinion on your startup idea. Free, evidence-backed, ready in 10 minutes.
npx skills add eph5xx/tweakideaWhat you get:
- A scorecard — verdict (PIVOT / GO / STOP), weighted score, letter-graded evidence
- A 14-dimension breakdown — per-dim scores with what's strong, what's weak, what's untested
- An assumption ledger — every untested claim in your idea, surfaced so you know what to validate next
- A shareable HTML report — opens in your browser, easy to revisit
Why not just ask ChatGPT?
- 14 evaluators run in parallel — one per dimension
- Each runs its own independent targeted web search and cites sources
- Every untested assumption in your idea is surfaced
- Output is a structured beautiful scorecard you can revisit and share, not a one-shot paragraph
Browse full interactive reports: Airlock · Mirror · Giftwrap
An idea dies from what you don't verify. TweakIdea runs an idea through 14 weighted dimensions — pain intensity, willingness to pay, solution gap, founder-market fit, and ten more — each scored with cited web research. Fourteen Sonnet evaluators work in parallel; Opus merges the results into a scorecard with strengths, weaknesses, and three rewrites targeting the weakest dimensions. The output names what you haven't tested yet. It doesn't tell you to quit your job.
Longer background: TweakIdea — the origin story · TweakIdea v2 — what changed
You: /tweakidea evaluate "A mobile app that lets restaurants sell unsold
food at a discount 30 minutes before closing"
Claude: [researcher kicks off in background — web search on market,
competitors, pricing]
[extracts hypotheses from your idea — e.g. "restaurants have
unsold food at close", "30-min window is enough to sell it"]
Mark each hypothesis: CONFIRMED / UNCONFIRMED / REJECTED?
You: [confirms what you already know; leaves the rest unconfirmed]
Claude: [2-4 founder-fit questions on your connection to THIS idea —
restaurant ops? distribution? capital?]
You: [answers inline]
Claude: [launches 14 parallel evaluators, one per dimension]
[each runs its own targeted web searches for evidence]
[compute.py aggregates per-dim scores, weights them, grades
evidence A+ through F]
[Opus synthesizes strengths/weaknesses, next steps, and a
potential-uplift narrative driven by your unconfirmed
hypotheses]
PIVOT — Promising, address weak areas | Weighted 3.4/5.0
| Potential 4.0/5.0 | Evidence C+
[14-dim scorecard; report.html opens in browser]
You: /tweakidea improve latest
Claude: [three rewrites — small reframe, medium reshape, big reimagine]
[each targets the weakest dimensions with rubric-grounded rationale]
[paste any rewrite into /tweakidea evaluate to re-score]
Runs 14 independent Sonnet evaluators in parallel — one per dimension — then merges them into a weighted scorecard with assumption tracking and letter-graded evidence. Produces a markdown scorecard and an HTML report.
/tweakidea evaluate "A mobile app that lets restaurants sell unsold food at a discount"— inline/tweakidea evaluate path/to/idea.md— from a file/tweakidea evaluate— fully interactive/tweakidea "A mobile app that..."— shortcut: any text that isn't a command name is evaluated as an idea
Searches Hacker News via Algolia and prints a ranked table of threads worth feeding into /tweakidea analyze-hn-post. Each hit is scored 1–5 on how likely it is to yield a real startup-idea seed. Read-only.
/tweakidea browse-hn llm agents week— topicllm agents, last 7 days/tweakidea browse-hn robotics— topic only; prompts for a time window/tweakidea browse-hn today— window only; browses everything recent/tweakidea browse-hn— fully interactive
Fetches a Hacker News post (article + full comment tree), identifies 3-6 technology shifts, and surfaces 1-3 product opportunities per shift. Confirmed opportunities are written up and can be fed into /tweakidea evaluate.
/tweakidea analyze-hn-post 43374458— by HN id/tweakidea analyze-hn-post https://news.ycombinator.com/item?id=43374458— by URL/tweakidea analyze-hn-post— prompts for an id
Reads a completed evaluation run and generates three rewrites at escalating scales — Small reframe, Medium reshape, Big reimagine — each targeting the weakest dimensions with rubric-grounded rationale and mandatory trade-off acknowledgments. Read-only.
/tweakidea improve latest— improve the most recent run/tweakidea improve 20260416— improve a run by timestamp prefix
Compares two evaluation runs side-by-side — score changes, potential shifts, verdict movement, and a per-dimension breakdown of what moved. Read-only.
/tweakidea diff latest— compare the two most recent runs/tweakidea diff 20260413 20260416— compare two runs by timestamp prefix/tweakidea diff 20260413 latest— compare a specific run against the latest
Lists artifacts under ~/.tweakidea/ — evaluation runs, HN analyses, and founder profiles. Arguments are free-form. Read-only.
/tweakidea list— all categories, 5 most recent per section/tweakidea list runs 20— the 20 most recent evaluation runs/tweakidea list best ideas— runs ranked by weighted score/tweakidea list hn— just HN analyses
Opens any artifact under ~/.tweakidea/ by timestamp, keyword, HN id, founder name, or natural query. Run reports open in the browser; HN and founder artifacts are inlined.
/tweakidea show latest— most recent evaluation run/tweakidea show 20260412-143022— a specific run by timestamp/tweakidea show restaurant food waste— keyword match on idea text/tweakidea show hn 43374458— an HN analysis by id/tweakidea show best ideas— ranked query over saved runs
Uploads a run's report.html to a secret GitHub gist and prints two links — the gist itself and an htmlpreview.github.io rendered URL you can paste into Slack or email. Requires the GitHub CLI (gh auth login).
/tweakidea share— share the most recent run/tweakidea share latest— same, explicit/tweakidea share 20260412-143022— a specific run by timestamp/tweakidea share restaurant food waste— keyword match on idea text
# uv runs the bundled Python scripts (or: brew install uv)
curl -LsSf https://astral.sh/uv/install.sh | sh
npx skills add eph5xx/tweakideaWorks in any Agent Skills agent; tuned and tested on Claude Code. /tweakidea evaluate also needs subagents and web search.
This uses the skills CLI, which detects your installed agents and installs one skill, tweakidea. Useful flags:
-ginstalls for all projects instead of just the current one-a claude-codetargets a specific agent-yskips the prompts, for examplenpx skills add eph5xx/tweakidea -g -a claude-code -y
After install, open your agent and type /tweakidea followed by a command: evaluate, improve, diff, browse-hn, analyze-hn-post, list, show, or share. Run /tweakidea on its own to see the command menu.
Update with npx skills update tweakidea. Remove with npx skills remove tweakidea (add -g for a global install).
Note
First /tweakidea evaluate run takes 10-20 minutes — it creates your founder profile, runs the researcher, and dispatches 14 parallel Sonnet evaluators plus one Opus synthesis pass. Subsequent runs reuse the founder profile and finish faster.
Claude Code:
/plugin marketplace add eph5xx/tweakidea
/plugin install tweakidea@tweakidea
Plugin skills are namespaced, so commands become /tweakidea:tweakidea <command>.
Codex:
codex plugin marketplace add eph5xx/tweakidea
codex plugin add tweakidea@tweakideaStart a new session afterwards. You can also install it from /plugins inside Codex. Update with codex plugin marketplace upgrade tweakidea.
Cursor: TweakIdea isn't in the Cursor Marketplace, so add the repo as a marketplace:
cursor-agent plugin marketplace add https://github.com/eph5xx/tweakideaThen run cursor-agent, type /plugin, open the Marketplace tab, and select tweakidea.
Or clone it into Cursor's local plugin folder and run Developer: Reload Window:
git clone https://github.com/eph5xx/tweakidea ~/.cursor/plugins/local/tweakideaLegacy installer (npx tweakidea)
The original npm installer still works for Claude Code. It prompts for global (~/.claude) or local (./.claude) placement and copies the skill there.
npx tweakidea # install
npx tweakidea -u # uninstallBetter article extraction (optional)
For JS-heavy sites, run uv run playwright install chromium once. Without it, the HN fetcher falls back to plain HTTP — which works fine for most sites.
MIT — see LICENSE for details.