Lightweight, local-first skills for Claude Code.
Three tools designed for AI coding agents, with all data stored as simple files in your repo:
| Skill | What it does |
|---|---|
| issues | Local issue tracking inspired by GitHub Issues. A more granular companion for tracking work during development sessions. |
| sessions | Session memory that persists learnings, open questions, and next actions across conversations. Fully inspectable and user-managed. |
| adr | Architecture Decision Records using RFC/PEP-style format. |
All data lives in your repo as append-only JSONL files - no external services, no daemons, easy to read and edit directly.
Each skill can be used independently, but they work well together:
issues (standalone)
^
|-- sessions (can reference issues via issues_worked)
|-- adr (can link related issues)
- issues is fully standalone - copy it alone to any repo
- sessions optionally tracks which issues you worked on
- adr optionally links to related issues in decision records
# Clone and install
git clone https://github.com/davidcpage/skill-issues.git
cd skill-issues
uv tool install -e .
# Initialize in your project (installs all skills)
issues init --all /path/to/your/projectThis copies skill files and sets up permissions automatically.
To update: git pull (in the skill-issues directory)
You can also install skills individually: issues init, sessions init, adr init
Data directories (.issues/, .memory/, .decisions/) are created automatically on first use.
These skills are designed for natural conversation with Claude Code. Just describe what you want - Claude handles the tooling.
Starting a session:
"What issues are ready to work on?"
There's one issue ready to work on:
- #050 - Refocus README on skill usage with examples (task, priority 2)
Creating issues:
"Create an issue for the login timeout bug - it's high priority"
Created #051 - Login timeout bug as a high-priority bug.
During work:
"Add a note to issue 051 - discovered it only happens with expired sessions"
Added note to #051.
"Close issue 051 - fixed by refreshing auth tokens before API calls"
Closed #051 - the auth token refresh fix resolved the timeout.
Visualizing dependencies:
"Show me the issue dependency diagram"
flowchart LR
050["050: Refocus README..."]
051["051: Login timeout"] --> 050
style 050 fill:#87CEEB
Resuming context:
"What did we work on last session?"
Last session (s027, Dec 15) we worked on block/unblock commands and closed issue #045.
Key learnings:
- Post-hoc dependency modification is a reasonable escape hatch
- Two explicit commands are cleaner than one that replaces the whole list
Finding open threads:
"What open questions do we have?"
Open questions across recent sessions:
- How well does this scale with multiple concurrent agents?
- Should notes support optional session field for tracking context?
- What's the best repo structure for publishing skill bundles?
Ending a session:
"Let's wrap up - we learned that Mermaid LR layout works better for tall diagrams"
Created session s028 capturing learnings and issues worked.
When facing a design choice:
"We need to decide between SQLite and JSONL for storage - let's write up an ADR"
Claude creates a decision record in .decisions/ capturing context, options, and rationale:
.decisions/
├── 001-sessions-vs-issues.md
├── 002-github-issues-compatibility.md
└── 003-design-doc-process.md
The issues and sessions commands (installed during Quick Start) also work directly from the terminal:
Issues CLI:
issues # Open issues (default)
issues --ready # Open and not blocked
issues --closed # Closed issues
issues 053 # Show single issue
issues --create "Title" # Create new issue
issues --close ID "Reason" # Close issue
issues --diagram # Dependency diagram
issues board # Interactive Kanban TUISessions CLI:
sessions # Last session
sessions --last 3 # Last N sessions
sessions --open-questions # All open questions
sessions --create "topic" # Create session
sessions --amend -l "learning" # Add to last session
sessions board # Interactive TUI browserTUI Interfaces:
issues board - Kanban board with Ready/Blocked/Closed columns, vim navigation (h/l/j/k), details panel.
sessions board - Session browser with date list, search filter (/), full session details.
For full CLI documentation, see:
.claude/skills/issues/SKILL.md.claude/skills/sessions/SKILL.md.claude/skills/adr/SKILL.md
This repo was built using its own skills. The .issues/ and .memory/ directories contain the actual issues and session logs from development:
- 78 issues tracked from initial prototype to publishable skills
- 43 sessions capturing learnings about append-only logs, skill design, and more
- 6 ADRs documenting key design decisions
Explore them to see the skills in real use.
- Local-first - No external services, just files in your repo
- Git-friendly - Append-only JSONL means clean diffs and easy merges
- AI-native - Optimized for Claude Code, not human CLI ergonomics
- Lightweight - No daemons, databases, or complex setup
This project was inspired by beads by Steve Yegge - a sophisticated agent memory system with git-backed JSONL, SQLite caching, and daemon architecture. These skills are a simplified take on similar ideas.
While exploring why Claude Code can work fluently with tools like beads or these skills with no special training, Claude Opus 4.5 introduced the term protocol fitness to describe the phenomenon. AI agents have seen millions of issue trackers, RFCs, and changelog formats in training data. This means they already understand the workflows - how to triage issues, track blockers, close with a reason, link related work.
The development logs capture more discussion of this idea.
Issues and PRs welcome.
MIT