remind-agent

AI-assistant context for the REMIND model, packaged as a self-contained .claude/ bundle: a lean always-on CLAUDE.md plus four task skills that auto-load by task, their reference docs and scripts, and optional slash-command shortcuts.

What's in here

• CLAUDE.md — always-on base: HPC/safety rules, data hygiene (native-tools-first), and a compact REMIND orientation.
• skills/ — four task skills, each a SKILL.md + reference/ + (where useful) scripts/:
  • debug-remind-run — diagnose a failed/non-converged run (Cases A–F, non-convergence, TIRF).
  • inspect-remind-gdx — inspect a gdx / verify a gdx writer change.
  • configure-remind-scenarios — author/review scenario configs.
  • edit-remind-code — read/change GAMS/R code, compile-after-edit, pik-piam/renv.
  • write-handoff — only on request (/write-handoff <topic>): writes scratch/HANDOFF-<date>-<topic>.md so a later session can pick up the work.
• commands/ — optional /shortcuts that jump straight into a skill's playbook.
• scratch/ — working dir for agent temp files (contents gitignored).

Deployment

The repo is the .claude/ directory. From the root of a REMIND checkout:

git clone https://github.com/remindmodel/remind-agent.git .claude
printf '.claude/\n' >> .git/info/exclude     # keep it out of REMIND's git
rm -f CLAUDE.md                              # drop any stale root CLAUDE.md pointer, if present

That's the whole install — Claude Code discovers .claude/CLAUDE.md, .claude/skills/, and .claude/commands/ natively. No symlinks, no @-imports.

Optional: browse the bundle under a friendlier name with a cosmetic alias (tools still read the real .claude/, so it never sits in a discovery path):

ln -s .claude remind-agent

Do not do the reverse (ln -s remind-agent .claude): a symlinked .claude currently breaks skill invocation in Claude Code (/skill → Unknown skill).

Updating

git -C .claude pull

Portability (other agents)

The content is tool-neutral. opencode already reads .claude/skills/ natively, so the skills work there with no change — add a root AGENTS.md (or an opencode.json instructions entry pointing at .claude/CLAUDE.md) for its always-on memory. qwen-code reads AGENTS.md / QWEN.md the same way. The always-on file is kept as CLAUDE.md because it always loads; AGENTS.md is the portable alternative but only loads on Claude Code ≥2.1.277 and is suppressed by any CLAUDE.md above it.

Scope

REMIND-only. Cross-project, machine-specific skills (madrat development, general R dev) are personal — keep them in your own ~/.claude/skills/, not in this shared bundle.