What I use AI for
I work with coding agents to investigate problems, compare approaches, and build software. I like making complex things simpler. I want an agent to help me make a decision, not quietly make it for me.
Agents & harnesses
- Pi: My main coding harness, using my ChatGPT and Claude subscriptions rather than separate chat apps.
- claude-bridge: The extension I use to bring Claude into Pi.
Apps & context
- Wispr Flow: Occasional voice dictation; I mainly type.
- Herdr + Ghostty: The local workspace and terminal environment hosting this Pi session on macOS.
- Obsidian: Git-backed Markdown for plans, research, decisions, and handoffs so a fresh agent session can resume from an agreed checkpoint.
Skills & extensions
My main public customisations live in pi-skills. Skills give the agent task-specific guidance; extensions add tools and session UI. Many skills are adapted from other people's work, with attribution in the repository.
- Ship skills: Available guidance for alignment, research, design, and delivery planning when a change needs more than direct implementation.
- Coding and review skills: Focused guidance for debugging, architecture, language conventions, security, and usable interfaces.
- Delegate: Available in Pi for bounded background agent tasks, with an Agent Desk for inspection and control.
- Intercom: Available for local Pi sessions to inspect persisted context and exchange targeted messages; each session retains ownership of its work.
- PR monitor: Available to bring new GitHub review feedback back into the owning Pi session without treating feedback as permission to push or merge.
How I work
My standing instructions define a small-step workflow for substantial changes:
- Investigate before editing. If I ask why something behaves a certain way, read the relevant code, tests, docs, and history. A question is not permission to change files.
- Agree on the change. Compare credible approaches, explain the trade-offs, and confirm what to change and how to test it before implementation.
- Complete one coherent chunk. Implement and validate the approved change, preserve unrelated behavior, and stop at the agreed checkpoint. Keep the decision and evidence in the vault rather than only in chat.
- Check before delivery. Run focused checks during the work and the project's regression checks before finishing. Commits, pushes, and PRs require scoped approval; an unexpected failure is a reason to reassess, not expand the task.
Drafted with AI assistance.