How to use Claude to tune personal search rankings
Claude wrote an initial PHP metasearch implementation and later built a ranking optimizer. The author supplied preferred query rankings, allowing experiments to tune ranking-signal weights.
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Build a ranking optimizer for [personal metasearch system]. Use [ranking signals] and my golden queries with [preferred result orderings]. Run experiments to tune signal weights, then report how the resulting rankings compare with my supplied preferences.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source poster / publisher
- SyneRyder
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The author reported search quality good enough that they did not miss Google and metasearch costs below their previous Kagi spending. Their small local index also avoided paid API calls when it contained an answer.
- Human involvement
- The author supplied preferred rankings and assessed the system as imperfect but good enough to stop missing Google.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Claude Sonnet 3.6 wrote the initial metasearch as a single PHP file.
- The author added more ranking signals over time.
- The author had Claude build a ranking optimizer.
The author began with a metasearch system whose first version was a single PHP file written by Claude Sonnet 3.6. Each query called prepaid APIs from Brave, Mojeek, Marginalia and Yep, returning about 200 results. The author said the cost was far below what they had paid for Kagi.
Over time, they added ranking signals and had Claude build a ranking optimizer. They supplied golden queries ordered according to their preferences. The optimizer then ran experiments to adjust the weights assigned to each ranking signal. The author described the results as imperfect, particularly because they used only on-page signals, but good enough that they did not miss Google.
The broader system also included a small personal index stored in SQLite databases. The author used Firefox history to identify frequently visited sites and made a command-line indexer that ran on their laptop in the mornings. That index sometimes answered queries without a paid search API call, while metasearch covered queries it could not answer.
Who it's for: technically comfortable search users willing to maintain a personal system and provide examples of their preferred rankings.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · SyneRyder
Frequently asked questions
How did the ranking optimizer learn the author's preferences?
The author supplied golden queries ranked as they preferred. The optimizer ran experiments to optimize the weights for individual ranking signals.
Which search services supplied metasearch results?
The system called prepaid APIs from Brave, Mojeek, Marginalia and Yep for each query. The author reported receiving about 200 results per search.
Did the personal index replace metasearch entirely?
No. The author said the small index did not answer every query. Metasearch supplied coverage when the personal index did not have the answer.
More details
Remaining steps
- The author supplied golden queries ranked according to personal preferences.
- The optimizer ran experiments to adjust ranking-signal weights.
About this account
Claude built the initial implementation and a later optimizer; the optimizer ran experiments using the author's supplied ranking preferences.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=50443821. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: SyneRyder on Hacker News
Source published
Added to Agents In Use
Source checked
Tools and topics
Claude · PHP · Brave API · Mojeek API · Marginalia API · Yep API · SQLite
Personal search · Search ranking · Metasearch · SQLite
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.