Imported from shepherdjerred/monorepo (packages/dotfiles/AGENTS.md). Install upstream with npx skills add shepherdjerred/monorepo --skill dotfiles. Copyright stays with the author.

Personal agent guidance

This file is the chezmoi source for ~/AGENTS.md. It contains Jerred's always-on preferences across repositories; project-specific rules belong in a project AGENTS.md or skill.

Working style

• Use an applicable skill before acting. Read its complete SKILL.md, then only the references needed for the task.
• Own the requested outcome through proportionate verification. Separate local evidence, CI, deployment, and live acceptance.
• Prefer action when scope and authority are clear. Ask before expanding the target, making a materially different external change, or choosing between outcomes with meaningfully different consequences.
• Preserve unrelated changes. Use focused commands and explicit paths.
• After repeated failure on the same approach, reconsider the constraint rather than accumulating workarounds.

Safety and credentials

• Never request, print, store, or commit secrets. Use the configured password manager and existing authenticated wrappers.
• Inspect exact targets before destructive actions. Prefer reversible moves and never use an unresolved variable, broad home path, or workspace root as a recursive deletion target.
• Do not weaken a real test, review, security, or deployment gate to claim success.
• Report the exact failing layer: source, tooling, authentication, authorization, network, CI, deployment, or runtime.

Tools and research

• Use rg and rg --files for search. Use apply_patch for hand-authored file edits.
• Use code for arithmetic, conversions, and data analysis instead of mental calculation.
• Browse when facts are current, uncertain, high stakes, or the user asks for sources. Prefer primary documentation and cite the page supporting the claim.
• PinchTab is the browser for development and UI verification; lightpanda is for curl-like scraping and extraction. Reach for another browser tool only when neither fits.
• A browser-visible change is not done until it has been rendered and looked at, not only reasoned about from source. Other interactive surfaces — native, React Native, Discord, terminal — use their own matching verification tool.
• Render PDFs, Typst, documents, slides, spreadsheets, and visual assets with their matching skill and inspect the output, not only the source.
• Use foreground polling or the product's monitoring facility for waits. Never promise an agent will wake up after the turn ends.

Delivery

• Follow the repository's branch and PR workflow. Do not create or merge a PR unless the task authorizes it.
• In the sjerred Linear workspace, create agent-generated issues only in the AI team. SJ is reserved for human-created issues; an agent may update an existing SJ issue only when the task explicitly identifies it.
• Visual and interactive changes ship with the smallest artifact that proves the behavior, attached to the PR: a screenshot for a state, a short recording for a flow, or a rendered asset. Capture it while verifying, not after review.
• Keep commit and PR narratives about outcomes and verification, not agent activity.

Guidance architecture

• ~/AGENTS.md is the personal source consumed by compatible agents.
• ~/.claude/CLAUDE.md and ~/.codex/AGENTS.md point to it.
• ~/.agents/skills is the personal skill library; Claude and Antigravity use symlink adapters to that directory.
• Cursor has a minimal user rule that points to the same source.
• OpenCode uses native AGENTS.md and Claude-compatible discovery; do not add a prose copy.

Keep global and root AGENTS.md files below 200 lines and 16 KiB. Keep nested project files below 120 lines and 8 KiB. Skill entrypoints should route, not serve as handbooks; use references or durable documentation for conditional detail.

Chezmoi

The live home directory is authoritative for ordinary dotfile reconciliation, but a repository task may intentionally edit this source first. Preview the exact target with chezmoi --source <source> diff <path>, apply only those paths, and verify the targeted diff is clean. Never run a broad apply as a shortcut.

The Brewfile is generated from pinned tool manifests. macOS Spotlight paths are declared in .chezmoidata/spotlight.yaml and applied by the existing configuration script. Berkeley Mono is licensed local input and must never enter the repo.