Key Takeaways

  • Daniel Blum, a product manager at Melio, built an automated weekly loop that updates his Claude setup without manual prompt tuning.
  • Instead of training an AI assistant through endless correction prompts, Blum's system audits the silent delta between what Claude drafted and what he actually sent.
  • Every custom skill logs silent friction points and user edits into context files, surfacing top operational bottlenecks every Sunday.
  • An auditor skill called 'Improve' filters incoming AI hype and Twitter tips through a private Slack channel before adding features to his personal setup.
  • The engine behind this compound efficiency is Daniel Blum's 4-Part AI Self-Improvement Loop.

Daniel Blum's 4-Part AI Self-Improvement Loop

Part 1: Draft Gap Analysis

Claude checks previously generated draft messages that received no direct feedback, compares them against what was actually sent across communication channels, and updates writing preferences based on the delta.

Part 2: Recurring Task Skill Mining

Claude audits recurring workflows, multi-step actions, or tool handoffs over the prior week (e.g., design handoffs or prototype builds) and suggests converting them into structured, reusable skills.

Part 3: Telemetry & Friction Auditing

Skills and recurring tasks log user corrections, friction points, and explicit fix requests into context files. The loop aggregates the top frictions weekly and proposes targeted code/prompt updates to existing skills.

Part 4: External Tip Evaluation ('Improve' Skill)

External articles, posts, and AI tips shared via a dedicated Slack channel are audited critically against the current setup to evaluate feasibility and necessity before integrating new features.

When This Works (and When It Doesn't)

This framework runs automatically on a weekly scheduled basis to refine and maintain an individual's personal Claude and Cowork setup without requiring dedicated manual prompt engineering time. As Claire Vo observed during the interview, most operators waste time manually tweaking prompts every time an assistant fails, whereas Blum turns background telemetry into an automated maintenance cycle: “All of my skills and recurring tasks have lines in their files to collect feedback and friction from my interaction with them.”

This system works best for operators with high message volume and repetitive cross-functional handoffs, like product managers, founders, and engineering leads. If your daily work does not involve writing across multiple channels or repeating multi-step tool interactions, the loop will starve for data. Draft gap analysis also requires broad read access across your outgoing communication tools to compare drafts against finished emails and Slack posts. Without that integration, Part 1 cannot measure the delta.

What to Do With This

Set up Part 1 and Part 3 this week. First, create a dedicated scratchpad file in your AI workspace titled friction_log.md. Add a single system instruction to your top three custom AI skills: "Whenever I edit your output, reject a suggestion, or ask for a formatting correction, log the exact friction point and the tool name to friction_log.md."

Next, schedule a recurring task for Sunday evening. Have Claude read friction_log.md, identify the three most common friction patterns, and output updated prompt rules directly into your skill configuration files. You will stop repeating the same five style corrections every Monday morning.