Current:
New: This reel provides a concrete, structured system prompt framework (CLAUDE.md) for orchestrating AI coding assistants, specifically detailing subagent delegation, autonomous bug fixing, and self-improvement loops, which is not present in the broader AI Web Dev Workflow plan.
Current:
New: While the Cloudflare plan discusses agent infrastructure, this reel introduces a specific, documented protocol for inter-agent communication and task management within a *coding assistant* context, including explicit directives for 'Subagent Strategy' and 'Task Management protocol' that are workflow-level rather than infrastructure-level.
Current:
New: This reel focuses on improving the *quality and efficiency* of individual agent output through structured prompting (e.g., 'Plan Mode Default', 'Self-Improvement Loop'), contrasting with the dashboard/marketplace's focus on managing agents and skills at a higher, abstract level.
Implement Boris Cherny's CLAUDE.md framework to formalize self-improvement loops and enforce staff-engineer quality standards through structured prompting.
Merge these 6 workflow sections into ~/.claude/rules/standards.md, specifically adding the '3+ steps' plan mode trigger and the lessons.md self-improvement loop
Adopt 'Autonomous Bug Fixing' language into agent_loop.py prompts and add the 6-step Task Management protocol to plan JSON metadata for better audit trails
Formalize the .qa/ directory structure to match 'Capture Lessons' workflow - require lessons.md updates after any correction found during review queue processing
Implement 'Subagent Strategy' for complex tasks on VPS2 - use the dispatcher to spawn one-shot subagents for research/exploration while keeping main context clean
We should adopt the 'Self-Improvement Loop' into our existing .qa/ directory structure - we already capture lessons but don't mandate reviewing them at session start, which would prevent repeat mistakes on the ReelBot agent_loop
The '3+ steps = plan mode' rule is chef's kiss. we've been using '2+ files' as our heuristic but step-based is cleaner. what's your trigger for entering plan mode?
What it is: A comprehensive CLAUDE.md structure (system prompt/rules file) that enforces strict workflow patterns on Claude Code including mandatory planning, subagent delegation, continuous learning from corrections, and autonomous execution standards.
How it helps us: Directly applicable to our existing 'Claude Upgrades' project. We already have rules consolidation (~1.4k tokens) and context monitoring, but lack the formalized 'Self-Improvement Loop' with lessons.md and the structured Task Management protocol. The 'Autonomous Bug Fixing' section aligns with our QA system but could be formalized.
Limitations: Subagent Strategy requires Claude Code Max plan (which we have but costs 5x per request) - not economical for every task. Some elements may conflict with our existing context-handoff skill at 75% threshold. The 'Verification Before Done' section assumes CI/CD infrastructure not all projects have yet.
Who should see this: Technical leads and developers using Claude Code for production work (Dylan/me). Specifically relevant for the claude-dispatcher project and ReelBot agent_loop.
| Step | Prompt | Completion | Cost |
|---|---|---|---|
| analysis | 19,282 | 2,519 | $0.0142 |
| similarity | 1,493 | 600 | $0.0006 |
| plan | 11,133 | 4,452 | $0.0148 |
| Total | $0.0296 | ||