Claude-Multi-Agent-Research-System-Skill
CLAUDE.md11 starsChanged 11 months agoApache-2.0View on GitHub
@import ~/.claude/CLAUDE.md For all prompts, as a first and final hook, remember to include a date/time stamp your response, as well as the execution time from start to finish. @import ../docs/workflows/research-workflow.md @import ../docs/workflows/planning-workflow.md @import ../docs/workflows/compound-request-handling.md @import ../docs/workflows/semantic-search-hierarchy.md @import ../docs/configuration/configuration-guide.md @import ../docs/guides/token-savings-guide.md This project uses 3 specialized skills with hook-based auto-activation. The imported workflows above contain detailed trigger keywords, examples, and step-by-step instructions. These critical rules are high-level decision gates to ensure you use the orchestration system correctly. ALWAYS use when:…
In this repository3 other files it gives its agents.
The file
We split this CLAUDE.md into the rules it gives an agent
11 rules. About 0 are common, 10 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈10
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 34
Use WebSearch/WebFetch directly for research tasks
- Distinctive: line 35
Write synthesis reports yourself (Write tool excluded when skill active)
- Distinctive: line 39
Verify all research notes exist before synthesis
- Distinctive: line 52
Start planning without invoking skill first
- Distinctive: line 68
Use Grep/Glob as first attempt for functionality searches
- Distinctive: line 81
MUST: use AskUserQuestion with standard options
- Distinctive: line 82
Wait for user response before invoking any skill
- Distinctive: line 134
NEVER calculate elapsed time mentally: - Always use Python code
- Distinctive: line 135
ALWAYS show both UTC and local: when reporting times
- Very specific: line 136
ALWAYS label timezone: - Never write time without UTC/CET/local annotation
- Distinctive: line 151
Decision: Use direct bash scripts for automatic reindex (session start, post-write hooks)
≈ “Common” is an estimate: at least 1 in 5 of the 27,641 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Project Instructions for anthropic_research
@import ~/.claude/CLAUDE.md
CRITICAL: Response Timestamp Requirement
For all prompts, as a first and final hook, remember to include a date/time stamp your response, as well as the execution time from start to finish.
@import ../docs/workflows/research-workflow.md @import ../docs/workflows/planning-workflow.md @import ../docs/workflows/compound-request-handling.md @import ../docs/workflows/semantic-search-hierarchy.md @import ../docs/configuration/configuration-guide.md @import ../docs/guides/token-savings-guide.md
CRITICAL: Universal Orchestration Rules
This project uses 3 specialized skills with hook-based auto-activation. The imported workflows above contain detailed trigger keywords, examples, and step-by-step instructions. These critical rules are high-level decision gates to ensure you use the orchestration system correctly.
Research Tasks: multi-agent-researcher Skill
ALWAYS use when:
- Multi-source research (2+ sources)
- Topic investigation requiring synthesis
- Comparative analysis, literature reviews, market research
NEVER:
- Use WebSearch/WebFetch directly for research tasks
- Write synthesis reports yourself (Write tool excluded when skill active)
MANDATORY:
- Delegate synthesis to report-writer agent (reads research_notes/, writes to reports/)
- Verify all research notes exist before synthesis
Self-check: Multi-source? Synthesis needed? >3 searches? → Use skill
Planning Tasks: spec-workflow-orchestrator Skill
ALWAYS use when:
- New features or systems (non-trivial additions)
- Architecture changes, multi-component work (>3 files)
- Requirements gathering, formal specifications, ADRs
NEVER:
- Manual planning with TodoWrite for non-trivial work
- Start planning without invoking skill first
MANDATORY:
- Quality gate enforced: 85% threshold (100 points, 4 criteria, 3 max iterations)
- 3-agent workflow: spec-analyst → spec-architect → spec-planner
Self-check: New feature? Specs needed? >3 files? → Use skill
Content Search: semantic-search Skill
ALWAYS use FIRST when:
- Searching for functionality/content by describing WHAT it does
- Finding patterns, similar implementations, documentation
- Exploring unfamiliar codebase
NEVER:
- Use Grep/Glob as first attempt for functionality searches
- Skip semantic search and go directly to keyword searches
VALUE PROPOSITION:
- Saves 5,000-10,000 tokens vs traditional Grep exploration (~90% reduction)
- Example: "user authentication logic" → 1 search, 2 file reads vs 15+ Grep attempts, 26 file reads
Self-check: Searching for functionality? → Semantic search first, Grep as fallback
Compound Requests
When BOTH research AND planning triggered simultaneously:
- Hook message: "COMPOUND REQUEST DETECTED"
- MUST use AskUserQuestion with standard options
- Wait for user response before invoking any skill
- NEVER assume user intent or skip clarification
System Architecture
Hooks (user-prompt-submit.py): Auto-detect trigger keywords → suggest/activate skills
Skills: Orchestrate specialized agents in parallel/sequential workflows
Agents: Execute research, planning, synthesis, semantic search operations
Progressive Disclosure: @import loads detailed workflows on-demand, keeping core context focused
For detailed workflows, trigger keywords, examples, configuration reference, and token economics, see imported documentation above.
CRITICAL: Timestamp Analysis Rules
MANDATORY for ALL timestamp operations to prevent timezone confusion errors.
NEVER Do Mental Math on Timestamps
❌ WRONG: "Last index 11:10, now 12:18, so ~68 minutes" ✅ CORRECT: Use utility script or code function
Three-Layer Prevention System
Layer 1: Use Utility Script for Ad-Hoc Analysis
python .claude/utils/verify_timestamp.py <file_path> <field_name>
python .claude/utils/verify_timestamp.py --raw "2025-12-11T11:10:21+00:00" "Event"Layer 2: Use Built-In Function for Code Analysis
from reindex_manager import get_reindex_timing_analysis
timing = get_reindex_timing_analysis(project_path)
print(f"Last reindex: {timing['last_reindex_utc']} ({timing['last_reindex_local']})")
print(f"Current time: {timing['current_utc']} ({timing['current_local']})")
print(f"Elapsed: {timing['elapsed_display']}")
print(f"Status: {timing['cooldown_status']}")Layer 3: Always Label Timezones Explicitly
- ✅ "11:10 UTC (12:10 CET)"
- ✅ "Current: 12:18 CET (11:18 UTC)"
- ❌ "11:10" (missing timezone)
- ❌ "now 12:18" (which timezone?)
Mandatory Rules
- NEVER calculate elapsed time mentally - Always use Python code
- ALWAYS show both UTC and local when reporting times
- ALWAYS label timezone - Never write time without UTC/CET/local annotation
- USE utility script for any ad-hoc timestamp analysis
- USE get_reindex_timing_analysis() for reindex timing in code
Why This Matters
Past Error: Reported "~68 minutes elapsed" when actual elapsed was 8 minutes Root Cause: Mixed UTC timestamp (11:10) with local time (12:18) in mental calculation Prevention: These mandatory rules + utility tools
Architecture Decision Records
Auto-Reindex Design: Direct Script vs Agent
- Decision: Use direct bash scripts for automatic reindex (session start, post-write hooks)
- Rationale: 5x faster (2.7s vs 14.6s), $0 cost vs $144/year per 10 developers, works offline, predictable
- Agent Use: Reserved for manual operations where intelligence and rich output add value
- Full ADR:
docs/architecture/ADR-001-direct-script-vs-agent-for-auto-reindex.md - Quick Reference:
docs/architecture/auto-reindex-design-quick-reference.md
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

