llm-collab

Agent MUST load when dealing with ADRs, devlogs, or multi-session documentation

Safety Notice

This listing is imported from skills.sh public index metadata. Review upstream SKILL.md and repository scripts before running.

Copy this and send it to your AI assistant to learn

Install skill "llm-collab" with this command: npx skills add bukzor/bukzor-agent-skills/bukzor-bukzor-agent-skills-llm-collab

--- # workaround: anthropics/claude-code#13005 setup: | All projects that depend on this skill should have as CLAUDE.md frontmatter:

```yaml
--- # workaround: anthropics/claude-code#13003
depends:
    - skills/llm-collab
```

LLM-Collaborative Documentation

Principles and helpers for human-LLM agent swarm collaboration on long-running projects. Provides documentation patterns for coordination across agents, temporal conflict resolution, and just-in-time context loading through structured files (ADRs, devlogs), progressive-detail guides, and optional scripts.

Note: This skill focuses on documentation patterns. For task management across sessions, use the llm-subtask skill which handles ephemeral, tactical, and strategic task tracking.

Status: Locally maintained, iteratively improving. Expect rough edges.

Core Principles

1. Separation of Concerns

Problem: Mixing audiences wastes everyone's time.

Pattern: Different files for different audiences/purposes:

project/
├── README.md                 # Users: what it does, how to use
├── HACKING.md                # Contributors: how to develop
├── CLAUDE.md                 # Agents: architecture, conventions
└── docs/dev/
    ├── adr/                  # Why: decision rationale
    ├── devlog/               # When: session history
    └── design/               # What/how: living design knowledge

2. Tiered Detail

Problem: Loading full docs when you need a summary wastes context tokens.

Pattern: Three-level information hierarchy:

  1. Top-level file: 1-2 sentence summary + link
  2. Main doc: 1-paragraph summary + link to deep dive
  3. Subdirectory: Full detail

Example:

CLAUDE.md: "Uses ADR format for decisions. See docs/adr/"
docs/adr/ directory: [Date-based ADRs, list with ls or grep]
references.kb/: [Categorized guides: file-types.kb/, guidelines.kb/, workflows.kb/]

3. Temporal Ordering

Problem: With 20+ sessions/day exploring ideas in parallel, codebase contains conflicting partial implementations. Which decision is current?

Pattern: Date-based naming provides temporal ordering:

  • ADRs: YYYY-MM-DD-NNN-title.md
  • Devlog: YYYY-MM-DD.md (or YYYY-MM-DD-HHMM.md)
  • When conflicts found, newest decision wins

4. Baton-Passing

Problem: Agent swarm loses alignment between sessions.

Pattern: Devlogs capture what diffs can't: reasoning, principles, conventions.

  • Decisions and their rationale (especially rejected alternatives)
  • Conventions established and principles discovered
  • Tradeoffs that shaped the approach

For task tracking: Use the subtask skill for managing ephemeral, tactical, and strategic tasks.

5. Living Documentation

Problem: Docs become stale and lie.

Pattern: Use directory listings as the source of truth:

  • docs/adr/ directory contains decisions—ls -t to see chronologically
  • docs/devlog/ directory contains session history—directory itself is the index
  • Avoid maintaining separate index files that can drift

6. Design Knowledge

Problem: CLAUDE.md is too brief for deep understanding; ADRs are too granular.

Pattern: Use docs/dev/design/ for living design documentation:

  • Deeper than CLAUDE.md, more distilled than ADRs
  • Organized by abstraction level (goals → requirements → design → components)
  • Linked via why[] frontmatter for traceability

When to read: Orienting on architecture, before significant changes.

When to update: After architectural ADRs, when onboarding reveals gaps.

See references/how-to-document-design-knowledge.md for the full pattern.

Frontmatter Directives

CLAUDE.md files use frontmatter to give agents operational instructions. These are action triggers, not passive metadata.

  • requires: — Read these files before acting in this directory.
  • depends: — Read when relevant.

Adaptation Guidelines

These are patterns, not rules. Adapt based on:

  • Team size: Solo vs multiple humans changes needs
  • Session frequency: 5 sessions/day vs 20 changes patterns
  • Exploration style: Serial vs parallel affects conflict frequency
  • Project phase: Early exploration vs maintenance
  • Existing conventions: Respect what team already uses

Evolution is expected. As you use these patterns:

  • Keep what works
  • Drop what doesn't
  • Modify to fit your needs
  • Share improvements

Quick Reference

Common workflows:

Key file types:

Common tasks:

# Orient yourself at session start
~/.claude/skills/llm-collab/bin/llm-collab-session-start

# Document a significant decision
~/.claude/skills/llm-collab/bin/llm-collab-adr "Decision title"

# Record what happened this session
~/.claude/skills/llm-collab/bin/llm-collab-devlog

# Backdate an entry (works with any date-based script)
DATE=2025-11-19 ~/.claude/skills/llm-collab/bin/llm-collab-adr "Decision title"

# Set up docs for new project
~/.claude/skills/llm-collab/bin/llm-collab-init

All scripts support --help for details.

Detailed References

Browse https:references.kb/ for categorized guides:

  • file-types.kb/ - What each doc type should contain
  • guidelines.kb/ - How to write effective docs
  • workflows.kb/ - How to use and maintain docs

Standalone references:

  • references/how-to-document-design-knowledge.md - Living design documentation pattern

Success Indicators

These patterns are working when:

  • Agents orient quickly (<1 minute)
  • "What's next?" has clear answer
  • Multiple daily sessions stay aligned
  • Conflicts resolve through temporal ordering
  • Context loading feels efficient
  • Both human and agents can drop in/out

Design Goals

These patterns address:

  • Projects with 10-20+ agent sessions per day
  • Parallel exploration creating conflicting partial implementations
  • Agents needing to hand off work to each other
  • User frequently asking "What's next?" or losing thread
  • Context waste from loading irrelevant documentation

Source Transparency

This detail page is rendered from real SKILL.md content. Trust labels are metadata-based hints, not a safety guarantee.

Related Skills

Related by shared tags or category signals.

Automation

claude-realignment

No summary provided by upstream source.

Repository SourceNeeds Review
Automation

llm-kb

No summary provided by upstream source.

Repository SourceNeeds Review
Coding

openclaw-version-monitor

监控 OpenClaw GitHub 版本更新,获取最新版本发布说明,翻译成中文, 并推送到 Telegram 和 Feishu。用于:(1) 定时检查版本更新 (2) 推送版本更新通知 (3) 生成中文版发布说明

Archived SourceRecently Updated
Coding

ask-claude

Delegate a task to Claude Code CLI and immediately report the result back in chat. Supports persistent sessions with full context memory. Safe execution: no data exfiltration, no external calls, file operations confined to workspace. Use when the user asks to run Claude, delegate a coding task, continue a previous Claude session, or any task benefiting from Claude Code's tools (file editing, code analysis, bash, etc.).

Archived SourceRecently Updated