Anybody else having trouble getting CC to put any details into issues?

[peterloron]

I'm using beads with Claude Code. In general I'm finding it quite helpful (thanks Steve!). However when I do something like:

Check this code for logic duplication and create issues for anything which needs to be looked at

Claude will create issues with a title, but no details. The issue is empty. I can point this out, and Claude will do the usual "You're right!" thing and fix it, but I'm wondering if anybody has some more strict/potent language for the AGENTS.md file...

[riordanpawley]

I'm using a skills/hooks system to get Claude to do that kind of thing right similar to

https://www.reddit.com/r/ClaudeAI/s/wrn2tjkMHX

[TheEditor]

If you've got a spin of your own on that would love to learn it.

[TheEditor]

Does your version of this hooks system work globally or do you have to install the guy's repo in every project?

[riordanpawley]

It's per-project (lives in .claude/ in each repo). The hooks/skills are checked into version control alongside your code.

To answer your setup question:

In your project root:

mkdir -p .claude/{agents,skills,hooks}
touch .claude/settings.json

Basic .claude/settings.json:

{
  "hooks": {
    "UserPromptSubmit": ".claude/hooks/user-prompt-submit.cjs"
  }
}

Basic hook (.claude/hooks/user-prompt-submit.cjs):
module.exports = async (prompt, context) => {
// Check if prompt matches skill triggers
// Return skill content to inject
return { skillsToLoad: [...] };
};

Key differences from the systems it's based on:

vs jefflester/claude-skills-supercharged:

• They use: Global AI pipeline, 2-skill slot limit, caching
• We use: Per-project hooks, no hard slot limits, project-specific triggers
• Same: <500 line rule, confidence thresholds, session deduplication

vs diet103/claude-code-infrastructure-showcase:

• They use: 2 core hooks (UserPromptSubmit, PostToolUse)
• We use: 4 hooks (added PreToolUse for blocking, Stop for cleanup)
• Same: skill-rules.json configuration, progressive disclosure pattern

Additional patterns we added:

• Agent/skill separation (agents reference skills, not duplicate them)
• Content pattern triggers (regex matching inside files, not just filenames)
• Anti-patterns (exclude false positives)
• Issue tracker integration (beads hooks for work tracking)

What they have that we don't:

• jefflester: Automated test suite (120 tests), /wrap sync command
• diet103: Generic examples (blog/posts domain), documented phase-based adoption

Practical setup steps:

1. Create .claude/skills/my-skill.skill.md with frontmatter:

---
skill: my-skill
confidence:
  required: 0.70
  suggested: 0.50
triggers:
  filePatterns: ["src/**/*.ts"]
  keywords: ["database", "query"]
---

2. Register in .claude/settings.json
3. Hook detects file/prompt patterns → loads skill

Trade-offs:

Per-project approach:

• ✅ Skills understand your specific codebase
• ✅ Version controlled with code
• ❌ Setup overhead per project
• ❌ Can't share generic patterns easily

You could make it semi-global by symlinking shared skills, but then you lose the codebase-specific knowledge.

Both upstream systems are solid foundations. Ours diverged to handle monorepo orchestration and work tracking, which you may or may not need.

---

TL;DR: Per-project setup in .claude/ directory. Based on diet103's hooks + jefflester's AI matching. Added agent orchestration and blocking hooks. Setup is manual.

[peterloron]

Interesting. I'll take a look. Thanks!

…

On Saturday, November 22nd, 2025 at 21:47, Riordan ***@***.***> wrote: I'm using a skills/hooks system to get Claude to do that kind of thing right similar to https://www.reddit.com/r/ClaudeAI/s/wrn2tjkMHX — Reply to this email directly, [view it on GitHub](#366 (comment)), or [unsubscribe](https://github.com/notifications/unsubscribe-auth/AABSMTNKBP3HMEKWEPPDNTD36FDABAVCNFSM6AAAAACM5VRZNSVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKMBVGAZDSMY). You are receiving this because you authored the thread.Message ID: ***@***.***>

[steveyegge]

Thanks for reporting this! We investigated and found this was indeed a real problem - even in our own projects we had significant rates of empty descriptions:

• beads project: 110/630 issues (17.5%)
• wyvern project: 8/119 issues (6.7%)
• vc project: 3/170 issues (1.8%)

We've shipped a two-part solution in commit bf74160:

1. Documentation improvements - Updated AGENTS.md, MCP tool docstrings, and Copilot instructions with:

• Prominent guidance section explaining why descriptions matter
• Good vs bad examples showing what to include (why/what/how)
• All code examples now include --description flags

2. Friendly validation warnings - When creating issues without descriptions:

⚠ Creating issue without description.
  Issues without descriptions lack context for future work.
  Consider adding --description="Why this issue exists and what needs to be done"

• Works in both CLI and daemon modes
• Smart enough to skip test issues
• Non-blocking (users can still proceed)

The warning approach is similar to how we handle test issues - gentle reminders rather than strict enforcement.

For the hooks/skills approach you mentioned, we've created bd-b0c8 to research that as a potential enhancement, but the documentation + warning combo should help significantly in the meantime.

Thanks again for the detailed report - this will help everyone using beads with Claude Code!

[TheEditor]

Do the improvements in bf74160 make it into the Claude Code skill?

[steveyegge]

(I'm going to leave this open until we see how well it works)

[peterloron]

I put this in CLAUDE.md just to see, and it definitely helps. Will switch to official once it releases.

CRITICAL: When creating or updating an issue, ALWAYS ALWAYS ALWAYS put details in the ticket describing the problem, completed troubleshooting, repro steps, etc. Anything which will be useful in maintaing context and not repeating work when the issue is worked on later.

[kevglynn]

What fixed this for me: workspace rules (.cursor/rules/) that fire on every message. One rule says "use --json for programmatic output" and another says "always include description, acceptance criteria for features, repro steps for bugs." The agent internalizes it without you repeating yourself.

The nuclear option that also works: a pre-commit hook on bd create that rejects empty descriptions. Agent learns within one bounce.