Autoresearch Growth Skill

此内容尚不支持你的语言。

Use the Autoresearch Growth skill when you want your coding agent to turn analytics data into review-ready growth experiments.

The inner loop generates, critiques, synthesizes, and blind-ranks candidates. The outer loop waits for real behavior, pulls the experiment data back into the next brief, and runs again from evidence.

This skill is based on Andrei Karpathy’s autoresearch pattern: keep the loop simple, let an external coding agent read the markdown instructions, write artifacts to disk, and iterate.

Agent Analytics adapts that pattern for growth work in the public template repo:

The important change is the outer evidence loop. Karpathy’s original loop is useful for generating and judging ideas. The growth version uses measured experiment behavior to seed the next run, so the loop does not stop at what LLM judges think sounds good.

Install

Install from the public skill repo:

npx skills add Agent-Analytics/agent-analytics-skill

If the installer asks which skill to install, choose agent-analytics-autoresearch.

You can also install it explicitly:

npx skills add Agent-Analytics/agent-analytics-skill --skill agent-analytics-autoresearch

The regular agent-analytics skill can be installed from the same repo. Use it first when the site still needs project setup, tracking, or basic experiment wiring.

What The Skill Needs

The skill works with any analytics source, but works best with Agent Analytics because the agent can collect the evidence itself.

Give the agent:

the target surface, such as homepage hero, pricing page, onboarding step, or signup CTA
the current control copy or screenshot
the product truth that must not drift
the audience
one primary event, such as signup, checkout, or activation_completed
one proxy event, such as cta_click
guardrails, such as bounce, scroll depth, time on page, errors, performance, source quality, or signup quality
recent analytics data or permission to fetch it

The skill can also run on demo data from the template repo, so you can try the workflow before connecting an account.

Run The Inner Loop

Start with a review-only run:

Run the autoresearch growth loop for this landing page. Use signup as the primary event and cta_click as the proxy. Fetch the latest analytics data if you can; otherwise use the data I provide. Keep product truth intact, run 5 rounds, and produce two distinct variants for review. Do not wire the experiment yet.

The skill should create or update the normal run artifacts:

brief.md
results.tsv
final_variants.md
a dated data snapshot when live data is available

results.tsv is the round log. The final decision artifact is final_variants.md.

Run The Outer Loop

Only move into production after human approval.

Once you approve a variant, ask:

Implement the approved variants on the landing page, create an experiment with control plus the two candidates, verify signup and cta_click tracking, and show me the QA links before launch.

After the experiment has collected behavior, ask:

Pull the experiment results and guardrails into a new snapshot. Summarize what won, what lost, what is still too sparse to trust, and run the next autoresearch loop from that evidence.

That is the loop on top of the loop:

run several LLM judging rounds
pick two variants for human review
ship an approved experiment
wait for real behavior
collect the measured result
feed the next autoresearch run

LLM judges are useful for pressure-testing options. They are not the final judge. User behavior is the final judge.

Try It With Sample Data

Clone or open the public template repo:

git clone https://github.com/Agent-Analytics/autoresearch-growth.git
cd autoresearch-growth

Then ask your agent:

Use the demo SaaS data in examples/demo-saas. Run the autoresearch growth loop and show me final_variants.md so I can judge the two winners.

No Agent Analytics account is required for the demo data.