Your first research loop
Create your first loop
We all love deep research — let's take "Deep Research" and let AI loopLoopWhat makes Claude an agent and not a chatbot. Instead of one ask-and-answer turn, a loop runs Claude over and over: act, observe what changed, decide the next step, repeat. on the idea.
Before
Prompt:
Web search on AutoResearch Application in business
Run this to see the before!
After
Prompt:
Web search on "AutoResearch Application in business". Maintain ONE report at AR-applications.md — edit and reshape it as you learn, don't just append. After each pass, think of new search terms to fill gaps and find new directions, and iterate. Loop 5 times. (do not use the /loop tool)
Results
The loop will run 5 iterations and stop on its own.
Permanent Plugin
Prompt:
Clone: https://github.com/fjfok/Humboldt
Place it in ~/Documents/github/
Install this plugin using /plugin
Then: ask Claude how to use it and try it.
Stuck?Debrief — what should have happened
An AR-applications.md file in your folder, substantially longer and more useful than a single web search would have produced. Notice the agent generates new search angles you wouldn't have thought of — that's not magic, it's just persistence applied to "what's missing here?".
You also just learned something about your own work: where in your day are you doing single-shot searches and stopping at the first decent result? That's a loop candidate.
Iteration beats intelligence. Humans have always known this — drafts, rehearsals, revisions. We just never had a tireless agent to do it for us.
Stuck?FAQ for this exercise
Q: Can you give a real case where this knowledge-distribution-via-md-file would work? A: Illustrative Scenario Imagine a conversational assistant like Zappia exploring how to improve suggested-questions prompts — analysed against history to promote feature exploration (because users who try ≥3 features get hooked). After a bounded number of iterations you'd have an .md summarising learnings on follow-up suggestions, which could transfer to other assistant patterns.
Q: Conceptual Example Consider a food-delivery contextContextEverything Claude can see right now — the conversation so far, the files it has read, the tool results, the system instructions. A big text buffer the model reads end-to-end before every reply. where ranking restaurants is an illustrative problem balancing customer experience and unit economics — could AutoResearchAutoResearchA loop that turns Claude into a tireless ML researcher. Give it a dataset and a metric to beat; it tries an approach, scores it, journals what it learned, and keeps going overnight. help find a better ranking algorithm? A: In principle yes — if you have a benchmark and a clear metric, no reason the pattern wouldn't apply. Internal teams have explored similar agentic workflows for ranking and propensity tasks in controlled experiments. Illustrative Scenario Pattern: small experiments → pick the best ones → scale up → A/B test in production.
Q: For ranking with new data coming in hourly — should we run back-to-back, or have a longer cadence? A: Define how often you want the agent to iterate based on when new data is available. For an offline benchmark, no need to wait — a shorter bounded iteration is fine. For live data updating hourly, match the cadence to the data refresh, iterating with checkpoints.
OptionalGo deeper
Re-run the loop with one constraint flipped — same topic, different shape.
Continue the loop on AR-applications.md, but this time:
- only accept findings backed by a primary source (paper, company report, ≥2024)
- skip blog posts and listicles
- when you find a contradiction with what's already in the file, mark it explicitly
Watch the search terms the agent invents change. A different constraint reshapes the path through the same problem — that's the lesson, not the topic.