AI Workflow Kit

Summary

AI Workflow Kit is a set of ready-to-use, markdown-defined skills that let an AI coding assistant run the recurring, non-coding parts of an engineer’s day: morning kickoffs, midday check-ins, end-of-day reviews, async standups, meeting prep, journaling, and code review. The skills are tool-agnostic — the same definitions work in Claude Code, Cursor, and Codex — and route through an integration layer instead of hard-coding any single service.

The problem

The biggest gap in most engineers’ productivity isn’t writing code — it’s everything around it: context gathering before work, notes during meetings, the status update the team needs, the end-of-day review that sets up tomorrow. Each takes ~15 minutes and gets perpetually deferred.

Why it mattered

These tasks compound. Skipping them quietly erodes focus, team awareness, and continuity across days. Automating them where the engineer already works — the terminal / editor — is the only place the automation actually sticks.

Constraints

• Must work across multiple AI assistants without rewriting each skill.
• Must not lock a user into one notes/tasks/calendar/chat vendor.
• Setup had to be low-friction enough that people actually adopt it.
• Skills are prompts, not code — reliability has to come from structure, not a runtime.

Role

Creator and sole maintainer. Designed the skill format, the integration-abstraction approach, the setup wizard, and authored the skill library.

Architecture

Each skill is a markdown file following the Agent Skills specification — an emerging standard supported by Claude Code, Codex CLI, and VS Code Copilot. Skills reference integrations (“fetch my calendar events”) rather than concrete services; an integration layer resolves those to Google/Apple Calendar, Notion/Obsidian, Jira/Linear, Slack/Discord, etc. The recommended auth path is the Zapier MCP server, giving one connection to thousands of apps instead of per-service API keys.

Key decisions

• Markdown skills over code plugins — portable across tools, diffable, and editable by the user without a build step.
• Integration indirection — skills describe intent; the binding to a specific vendor happens at setup, so the same skill survives a tool switch.
• A setup wizard that asks which tools you use and wires everything, rather than shipping a config file to hand-edit.

Tradeoffs

• Prompts-as-skills are inherently less deterministic than compiled tools; the design leans on explicit structure and trigger phrases to compensate.
• Standardizing on the Agent Skills spec bets on an emerging standard — chosen deliberately for portability over a proprietary format.

Failure modes considered

• A skill firing against the wrong integration → mitigated by explicit setup-time binding.
• Assistant-specific quirks → mitigated by keeping skills declarative and testing across tools.

Outcome

A working, public kit (the author’s most-starred original repository) that covers workflow, dev, and personal skills, with a one-command setup. The companion article, “I Built 50 AI Skills for the Parts of My Day I Never Got Around To,” documents the broader system it grew from.

What I learned

Reliability for prompt-based automation lives in the surrounding structure — trigger phrases, clear step sequencing, and integration boundaries — far more than in clever wording.

What I’d change next

Extract the integration-abstraction layer into a documented, standalone spec others can build skills against, and formalize the cross-tool compatibility matrix.