Skip to content

agent-tooling.md

Latest commit

 

History

History
170 lines (114 loc) · 6.02 KB

agent-tooling.md

File metadata and controls

170 lines (114 loc) · 6.02 KB
Raw

AI agent tooling (Cursor, Claude Code, Codex)

Cross-tool setup for token-efficient, low-slop agent sessions on Room TBA.

Layer Tool What it fixes
Talk Caveman Verbose prose, recap paragraphs
Code Ponytail Over-engineering, mystery deps
Memory AGENTS.md + scoped .cursor/rules/*.mdc Re-explaining the repo every chat
Workflow .cursor/skills/* Load detail only when relevant

Repo ships Ponytail + agent contract rules. Caveman installs locally via bun run install:agent-tooling (skills land in .agents/skills/, gitignored).

Already in this repo (clone = done)

Path Role
AGENTS.md Index + policies
docs/agent-contract.md Human-readable contract
.cursor/rules/agent-contract.mdc Output + process (always on in Cursor)
.cursor/rules/ponytail.mdc YAGNI / smallest diff (always on in Cursor)
.cursor/rules/caveman-default.mdc Caveman full activation (needs install:agent-tooling)
.cursor/skills/agent-handoff/SKILL.md Epic handoff template
.cursor/skills/ponytail-review/SKILL.md Over-engineering review (Cursor; mirrors plugin skill)
.cursor/skills/hallmark/ Anti-AI-slop UI skill (Nutlope/hallmark); Room TBA overlay in references/room-tba.md
.cursor/skills/impeccable/ UI polish / critique / hooks on UI edits
.cursor/rules/hallmark.mdc Hallmark gates on Svelte/CSS/Astro UI (scoped)
CLAUDE.md Pointer for Claude Code / Codex

One-shot setup (maintainers)

bun run install:agent-tooling   # Cursor: Caveman skills → .agents/skills/
bun run install:agent-plugins   # Claude Code + Codex: ponytail + caveman plugins (user scope)

Restart Cursor after the first command.

Shopped skills (aisles 1–4)

bun run install:shopped-skills
Aisle Skills Scope
1 Workflow systematic-debugging, verification-before-completion, using-git-worktrees, finishing-a-development-branch project
2 Stack playwright-best-practices, svelte-code-writer, svelte-core-bestpractices, web-design-guidelines project
3 Anti-slop fuck-slop, context-canary, grill-me, junior-to-senior global (~/.agents/skills/)
4 UI interface-kit project

Vendored in repo (no install): hallmark, impeccable under .cursor/skills/. Rule: .cursor/rules/hallmark.mdc (map chrome, side panel, landing).

Project skills install to .agents/skills/ (gitignored). Pin versions in committed skills-lock.json; teammates run npx skills experimental_install after install:agent-tooling to sync project skills from the lockfile.

Triggers: grill-me / junior-to-senior before big epics · fuck-slop on issue/PR prose · web-design-guidelines at 320px map QA · interface-kit only on UI sprints (pairs with Ponytail) · hallmark audit <path> before map chrome / sidebar redesigns · Impeccable hooks on direct UI file edits.

Caveman (local install per clone)

Compresses reply prose (~50–65% fewer output tokens). Code blocks stay normal.

Cursor

bun run install:agent-tooling
# or: npx skills add JuliusBrussee/caveman -a cursor

Skills install to .agents/skills/ (per-clone, gitignored). .cursor/rules/caveman-default.mdc activates full mode in-repo.

Restart Cursor after first install.

Level When
full Daily implement/debug (default)
lite PR descriptions, issues for humans
off Architecture debates needing nuance

Say stop caveman or normal mode to revert mid-session.

Claude Code

bun run install:agent-plugins
# or manually:
# /plugin marketplace add JuliusBrussee/caveman && /plugin install caveman@caveman

Or prefix prompts with /caveman full.

Codex

bun run install:agent-plugins
# installs ponytail@ponytail and caveman@caveman (user scope)

Repo CLAUDE.md + AGENTS.md cover instruction-tier when plugins are off.

Compress bloated context

If AGENTS.md or CLAUDE.md grows huge:

# In Claude Code with caveman plugin
/caveman-compress AGENTS.md

Ponytail (in repo + plugins)

Cursor: .cursor/rules/ponytail.mdc + .cursor/skills/ponytail-review/SKILL.md — committed, no install.

Claude Code / Codex (hooks + /ponytail, /ponytail-review):

bun run install:agent-plugins

Manual Claude Code: /plugin marketplace add DietrichGebert/ponytail → /plugin install ponytail@ponytail

Manual Codex: codex plugin marketplace add https://github.com/DietrichGebert/ponytail.git → codex plugin add ponytail@ponytail

Review pass after a fat diff: @ponytail-review (plugins) or ask Cursor to use ponytail-review skill.

Token tactics

Input

  • Scoped rules (alwaysApply: false + globs) — see map-layout.mdc, side-panel.mdc
  • New session + handoff skill instead of endless thread
  • Point at paths; don't paste 400-line dumps

Output

  • Caveman full for chat; normal grammar for commits and PR bodies
  • Agent contract bans recap menus and decorative diagrams

Process

  • Same-PR tests (AGENTS.md)
  • Agent runs gh / vercel — no dashboard handoffs

Intensity dial (quick reference)

Task Caveman Ponytail
Bug fix / feature full on (rules)
PR / issue for humans lite or off on
Security / prod / irreversible off on (don't skip validation)
Post-implementation review any @ponytail-review

Upstream

Bump vendored ponytail.mdc when upstream ladder changes materially. Bump Hallmark by replacing .cursor/skills/hallmark/ from upstream skills/hallmark/.