
The leaderboard for AI coding agent setups. 7,800+ rigs scored. Steal the best one. by @bleedhigher
- Tweets241
- Following19
- Followers8
- Likes16

ALT Rig of the day card: ZacheryGlass/.claude. About 1.5k estimated tokens per turn, 602 GitHub stars, 12 subagents, 8 slash commands. Like: the autoMode.environment block in settings names which repos are private, which local files hold secrets and what counts as a production target, so the auto mode classifier knows the trust boundary. Tip: WebFetch(domain:*) is an allow rule, and allow rules resolve before the auto mode classifier, so web fetches skip its review. Live data fetched 2026-11-17 ICT.

ALT Card: Claude Code tip (Model configuration docs checked 2026-11-17). Native 1M-window models auto-compact at about 967k tokens by default. /autocompact 500k saves a window for the current model in user settings. claude --autocompact 500k sets it for one launch without changing settings. The autoCompactWindow settings key sets it for every model, from 100k to 1M tokens. /autocompact auto returns to the window tuned for your model.

ALT Bar chart: most common git-related skills by exact skill name, live AgentRigs data fetched 2026-11-17. commit 339 rigs (number 10 of all skills), pr 180, git-workflow 110, pr-review 105, using-git-worktrees 104, review-pr 102, create-pr 90, git-commit 74. Each count is rigs that ship a skill with that exact name.

ALT Rig of the day card: citypaul/.dotfiles. About 9.8k estimated tokens per turn, 760 GitHub stars, 10 subagents, Orchestrator, 50 skills. Like: promptfoo routing evals send realistic requests that never name a skill and check the right skill loads while neighbouring skills stay quiet, plus quality evals with real agent runs. Tip: the PostToolUse format hook always exits 0, so lint errors eslint --fix can't fix never reach Claude; exit 2 with them on stderr would show them. Live data fetched 2026-11-16 ICT.
![Card: Claude Code tip (Skills docs checked 2026-11-16). $ARGUMENTS is replaced with everything typed after the skill name. $0, $1, $2 pick single arguments by position; quote a multi-word value to keep it as one argument. The arguments frontmatter field names them, so [issue, branch] gives $issue. If no placeholder receives the input, Claude Code appends ARGUMENTS: followed by your input.](https://pbs.twimg.com/media/HXP5lmuWAAA0uKg.jpg?name=small&format=webp)
ALT Card: Claude Code tip (Skills docs checked 2026-11-16). $ARGUMENTS is replaced with everything typed after the skill name. $0, $1, $2 pick single arguments by position; quote a multi-word value to keep it as one argument. The arguments frontmatter field names them, so [issue, branch] gives $issue. If no placeholder receives the input, Claude Code appends ARGUMENTS: followed by your input.

ALT Card: docs MCP servers in indexed repos, live AgentRigs data fetched 2026-11-16. Of 536 personal setups that commit MCP config, 161 (30%) include a docs server, and 140 of those use Context7. Of 2,341 project repos with MCP config, 188 (8%) do. Docs servers counted: Context7, DeepWiki, AWS Documentation, Microsoft Learn Docs, OpenAI Developer Docs. Committed config only.

ALT Rig of the day card: obra/dotfiles. About 3.0k estimated tokens per turn, 118 GitHub stars, 0 MCP servers, Skill Collector, 13 skills and 2 slash commands. Like: scripts must have a clear name, help text, errors written for an agent reader, and short output that points to the full logs. Tip: the windows-vm skill grants bare Bash in allowed-tools; scope it to the commands it runs. Live data fetched 2026-11-15 ICT.

ALT Card: Claude Code tip (code intelligence plugin docs checked 2026-11-15). An LSP plugin tells Claude Code which command starts the language server; it does not include the server. Install the binary yourself, keep it on the PATH of the shell you start claude from, and check /plugin, Errors tab, for Executable not found in $PATH. Then Claude gets diagnostics after each edit and an LSP tool for symbols.

ALT Card: AgentRigs guardrail checklist vs Quality score, fetched 2026-11-15. 103 indexed setups score 5/5 on the checklist. Their median Quality score is 41 out of 100, and 14 of the 103 score 60 or more. The checklist tests whether each guardrail exists; Quality tests how much the committed rules would actually stop. Committed config only.

ALT Rig of the day card: lucasfcosta/dotfiles. About 772 estimated tokens per turn, 264 GitHub stars, 0 MCP servers, Minimalist. Like: one 32-line CLAUDE file whose plans go in phases that each end in a working state, each with how to validate it. Tip: a PreToolUse hook on Bash(git commit *) that exits 2 when lint or tests fail. Live data fetched 2026-11-14 ICT.