How to use Claude to keep coding context clean
An author built separate-context research, Git, and Bash sub-agents that return condensed findings instead of full files, raw diffs, and logs, reporting a 90–95% token reduction on heavy operations.
news.ycombinator.com
Prompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Clone https://github.com/houseworthe/house-agents . Detect my OS and copy the .claude directory to my current project. Verify all three agent files exist in .claude/agents/ directory. List available sub-agents to confirm they load. Test house-research by finding TODO comments.
Reach people exploring AI agents
See placements →What happened
The author reports a 90–95% token reduction on heavy operations and says the main instance remained focused on implementation.
Reported by houseworth on Hacker NewsHow the agent did it
- Built a research sub-agent returning findings with file and line references.
- Built a Git sub-agent returning impact summaries grouped by severity.
- Built a Bash sub-agent extracting key command results and errors.
Where a person stepped in
The author tested the three implemented agents; the post does not describe detailed review criteria for their summaries.
A person approved each step before the agent continued.
The author built three specialized Claude sub-agents to reduce the material entering the main coding conversation. Their problem was that codebase searches and large Git diffs filled the main context with thousands of tokens that the coordinating instance rarely needed. After several such operations, the author found there was little room left for implementation work.
The sub-agents ran in separate contexts and returned condensed summaries. The research agent searched codebases and documentation, reporting findings with file and line references rather than full file contents. The Git agent analyzed diffs and commits and grouped impact summaries by severity. The Bash agent ran commands and extracted useful information such as pass/fail counts and relevant errors from their output.
The author reported good testing results and a 90–95% token reduction on heavy operations, saying the main instance stayed focused on implementation. Three other agents, for MCP, screenshots, and data queries, remained unbuilt. The post also supplied an installation instruction covering cloning the repository, copying its Claude directory, verifying agent files, listing agents, and testing research with TODO comments. Who it's for: developers whose coding conversations repeatedly fill with search results, diffs, or command logs.
Prompt
Original prompt
Source: Hacker News · houseworth
Frequently asked questions
What did the three implemented sub-agents return?
The research agent returned condensed findings with file:line references; the Git agent returned impact summaries grouped by severity; the Bash agent extracted key command results and errors.
Were all six proposed sub-agents implemented?
No. The author had built research, Git, and Bash sub-agents. MCP, screenshot, and data-query agents were listed as ideas not yet implemented.
More details
Remaining steps
- Tested the sub-agents and reported token reductions on heavy operations.
About this account
The specialized sub-agents performed research, Git analysis and command processing in their own contexts and returned summaries.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=45587911. Prompt: Original (quoted from source). Sourced by the Agents In Use pipeline.
Credit: houseworth on Hacker News
Source published 15 Oct 2025
Source checked 14 Oct 2026
Tools and topics
Claude · house-research · house-git · house-bash
Context management · Subagents · Code research
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.