chore: agent-agnostic standards, workflows, and skill optimization#2366
chore: agent-agnostic standards, workflows, and skill optimization#2366TC-MO wants to merge 5 commits intochore/strengthen-vale-toolingfrom
Conversation
Move .claude/rules/ to standards/ at repo root so all agents have parity. Replace verbose style sections in AGENTS.md with condensed summary + pointers. Trim standards where Vale now enforces the check. Add information ordering principle. Update symlinks, supporting files, and lint ignores. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Move workflow processes and references from .claude/skills/ to workflows/ at repo root for agent parity. Rewrite all 4 SKILL.md files as summary + pointer adapters with AgentSkills spec compliance. Delete Cursor symlinks, rewrite .mdc files as thin pointers. Add check-frontmatter.sh script and workflows section to AGENTS.md. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Eval showed Sonnet hybrid (summary + pointer adapter with delegated standards subagents) is the best tradeoff: ~80% cheaper than old Opus approach, 4% faster, and catches more issues. Deterministic tools run in main process (no sandbox issues), subagents handle standards review. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
Preview for this PR was built for commit |
| description: Create structured tutorials for Apify Academy or Platform documentation. Use when user says "create a tutorial", "write a tutorial", "build a step-by-step guide", "convert this guide into a tutorial", or "tutorial for [topic]". Handles tutorial structure, learning objectives, prerequisites, step-by-step instructions, code examples, and troubleshooting sections. | ||
| argument-hint: topic | ||
| allowed-tools: Read, Write, Edit, Bash, Glob, Grep | ||
| allowed-tools: Read Write Edit Bash Glob Grep |
There was a problem hiding this comment.
Why not use comma separated values?
There was a problem hiding this comment.
to be spec compliant with agent skill spec. Claude accepts both so to make them more agent-agnostic we switch to spaces
.claude/instructions.md
Outdated
There was a problem hiding this comment.
While we're at it, why not axe this file comletely? It's only mentioned in the claude.json file, but that is a halucinated file that is not loaded as standard.
There was a problem hiding this comment.
yup good catch, both claude.json and instructions.md to be axed
.claude/skills/api-doc/SKILL.md
Outdated
| argument-hint: endpoint-path | ||
| allowed-tools: Read, Write, Edit, Bash, Glob, Grep | ||
| allowed-tools: Read Write Edit Bash Glob Grep | ||
| metadata: |
There was a problem hiding this comment.
I didn't try this yet, but is this nesting supposed to be there?
There was a problem hiding this comment.
nice, I think this might have to be reverted. Agent Skills spec nests argument-hint within metadata, but Claude Code ignores metadata completely
There was a problem hiding this comment.
honestly Anthropic not being compliant with an open standard trips me for I think 3rd time, where I just assume that they are, but in reality they aren't :D
There was a problem hiding this comment.
Even funnier is they actually are responsible for the agent-spec yet don't follow it
| argument-hint: file-path | ||
| allowed-tools: Read, Bash, Glob, Grep, Agent | ||
| allowed-tools: Read Bash Glob Grep Agent | ||
| model: sonnet |
There was a problem hiding this comment.
Why tho? I wouldn't force a model here. Sonnet is often not capable enough to catch many issues.
There was a problem hiding this comment.
Did you actually went through the writeup in notion? I tested this approach with multiple models & sub-agents and no sub-agents. Sonnet was the most competent of the bunch in actually finding the issues. Opus while good for generation missed a lot and is prone to overthinking.
There was a problem hiding this comment.
I just quickly want through the PR, will check notion
Move argument-hint out of metadata: block back to top-level - Claude Code doesn't recognize metadata: nesting. Delete claude.json (not a real Claude Code config) and instructions.md (only referenced by claude.json). PR review feedback from marcel-rbro. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
Preview for this PR was built for commit |
…rkflows Skills now live in `.agents/skills/` (the AgentSkills open standard path), discoverable by Codex, Gemini CLI, OpenCode, Cursor, and other compatible agents. Claude Code discovers them via symlinks in `.claude/skills/`. Workflow content (process.md, references/, scripts/) absorbed into each skill's own `references/` and `scripts/` directories per the spec structure. The `workflows/` directory is deleted. Standards remain at repo root (shared). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
Preview for this PR was built for commit |
3 participants
Draft - writeup in notion
Summary
.claude/rules/→standards/at repo root (agent-agnostic).claude/skills/→workflows/at repo rootDepends on #2365
🤖 Generated with Claude Code