Releases·Tool Release
We tested giving Claude Code a 450-line static prompt vs a self-learning harness (and measured a 38.4% drop in regression errors)
A Reddit post by a developer describes testing a 450-line static CLAUDE.md against a minimal 25-line root file with dynamic skill loading across 40 multi-file refactoring tasks on their backend repo. The dynamic setup cut regression errors 38.4%, dropped per-turn input tokens from ~18,500 to under 3,200, and reduced session token spend over 45%. The team open-sourced the tool as autoharness.
19 engagement1 sourceWed, Sep 23, 2026, 09:33 AM
AI Overview
- A team's 450-line CLAUDE.md caused Claude Code to ignore mid-file edge-case rules and follow stale, refactored file-path conventions
- In a 40-task benchmark, a 25-line root file with on-demand skill distillation cut input tokens per turn from ~18,500 to under 3,200 and session token spend by over 45%
- The open-sourced autoharness tool distills verified command sequences into SKILL.md skills, loads them on demand, and auto-prunes stale ones using pure Python stdlib
- The team identified stale, unpruned rules—not missing instructions—as the biggest drag on coding agent accuracy, actively training the model to fail CI
Entities
Related
Products·launched
autoharness — self-learning skill layer for Claude Code
autoharness is an MIT-licensed Claude Code plugin that automatically learns skills from the sessions you already run, instead of requiring hand-written, stale skill files. It targets developers using Claude Code on ongoing projects who repeatedly re-solve the same problems because nothing persists between sessions.
autoharness is an MIT-licensed Claude Code plugin that automatically learns skills from the sessions you already run, instead of requiring hand-written, stale skill files. It targets developers using Claude Code on ongoing projects who repeatedly re-solve the same problems because nothing persists between sessions.
3 engagement1 sourceredditFri, Aug 28, 2026, 07:23 AM
Social·Community
Open-source tool autoharness distills Claude Code sessions into skills to keep CLAUDE.md under 30 lines
A Reddit post argues that common Claude Code guidance — piling coding standards, architecture rules, test commands, and lint preferences into CLAUDE.md — degrades agent performance over weeks of use. The author proposes modular skill distillation as an automated alternative to maintaining one large static prompt.
A Reddit post argues that common Claude Code guidance — piling coding standards, architecture rules, test commands, and lint preferences into CLAUDE.md — degrades agent performance over weeks of use. The author proposes modular skill distillation as an automated alternative to maintaining one large static prompt.
4 engagement1 sourceredditThu, Sep 3, 2026, 05:04 PM
Social·Community
How do you stop Claude Code from undoing things your team already decided?
A developer with eight years of experience says their team uses Claude Code daily, but the agent repeatedly reverts code to patterns the team deliberately moved away from, because the reasoning sits in old PRs the agent never sees. The team tried encoding rules in CLAUDE.md, which works briefly, but the file grows unmanageable and many rules' origins are forgotten. The post asks how other teams handle institutional knowledge that Claude Code lacks.
A developer with eight years of experience says their team uses Claude Code daily, but the agent repeatedly reverts code to patterns the team deliberately moved away from, because the reasoning sits in old PRs the agent never sees. The team tried encoding rules in CLAUDE.md, which works briefly, but the file grows unmanageable and many rules' origins are forgotten. The post asks how other teams handle institutional knowledge that Claude Code lacks.
64 engagement1 sourceredditSun, Oct 4, 2026, 06:43 AM
Social·Community
Developer seeks workflow advice for first large professional project using Claude Code on backend and frontend
A developer preparing to use Claude Code for the first time on a large professional backend-and-frontend project asks the community for production-grade workflow advice. The post seeks practices for staying in control — spec-to-test processes, documentation upkeep, and guarding against regressions — rather than fully autonomous setups, reflecting growing use of AI coding agents in serious business applications.
A developer preparing to use Claude Code for the first time on a large professional backend-and-frontend project asks the community for production-grade workflow advice. The post seeks practices for staying in control — spec-to-test processes, documentation upkeep, and guarding against regressions — rather than fully autonomous setups, reflecting growing use of AI coding agents in serious business applications.
22 engagement1 sourceredditFri, Oct 9, 2026, 12:49 PM
Social·Community
I measured what actually reaches Claude Code's context window over 61 days of my own transcripts
A solo developer analyzed 61 days of Claude Code transcripts (190 sessions, 918 subagent files) to measure what actually reaches the context window, then released the resulting open-source memory orchestration setup, Hipocampo. Built entirely on native features (hooks, MEMORY.md, CLAUDE.md, skills, subagents, compaction) with no MCP, it organizes memory into resident, paged, and interrupt regimes. The measurements expose gaps: zero injected recalls in 190 conversations, hook outputs silently truncated to ~2 KB previews, and aggressive compaction removing up to 96.6% of context.
A solo developer analyzed 61 days of Claude Code transcripts (190 sessions, 918 subagent files) to measure what actually reaches the context window, then released the resulting open-source memory orchestration setup, Hipocampo. Built entirely on native features (hooks, MEMORY.md, CLAUDE.md, skills, subagents, compaction) with no MCP, it organizes memory into resident, paged, and interrupt regimes. The measurements expose gaps: zero injected recalls in 190 conversations, hook outputs silently truncated to ~2 KB previews, and aggressive compaction removing up to 96.6% of context.
5 engagement1 sourceredditWed, Sep 30, 2026, 11:12 AM