Self Improving Claude Code Sessions

原始來源與檔名：2026-08-11T094455+0800-Self Improving Claude Code Sessions.md

SOURCE | 資訊源評估

• 作者：Alex Dunlop
• 來源：Medium
• 核心主題：Claude Code, Workflow Optimization, Self-Improving Feedback Loops

NAPKIN | 餐巾紙

餐巾紙公式

Strong AI Implementation + Weak Human Scoping = Broken Workflow Structured Workflow Pipeline + TSV Feedback Log = Continuous Process Improvement

一句話

By implementing a rigorous, multi-stage workflow pipeline for Claude (from scoping to self-verification) and embedding a structured, tab-separated feedback log to track and consolidate repeated corrections into rules, a developer can create continuously self-improving AI coding sessions.

餐巾紙草圖

  +---------+    +---------+    +----------------+
  | Scoping |--->| Kickoff |--->| Implementation |
  | (Grill) |    | (Plan)  |    | (Claude Code)  |
  +---------+    +---------+    +-------+--------+
                                        |
  +------------------+          +-------v--------+
  |  Self-Verify     |<---------| Feedback Log   |
  | (MCP/CheckRules) |          | (TSV: Date,    |
  +--------+---------+          |  Tag, Mistake, |
           |                    |  Correction)   |
  +--------v---------+          +----------------+
  | PR Review Stages |
  | (Draft -> Ready) |
  +------------------+

ROUND 1: SKELETON | 骨架掃描

• Survivorship bias analogy: reinforce what is weak, not just what is already strong.
• Pre-AI process: strong QA/review, weak scoping/kickoff.
• Post-AI process initially: Claude is so good at implementation that everything else (scoping, review) feels weak.
• The New Workflow: Scoping (grilling), Kickoff (planning), Implementation, Self-verification (MCPs + critical rules), Draft PR/Review, Ready PR/Review, Feedback/Learning.
• The Lunchtime Idea: Adding a formal /feedback and /learn mechanism.
• Feedback Log format: TSV file tracking date, tag, mistake, correction. Once a tag hits 3 entries, /learn proposes a durable rule.
• Why not just use Claude’s built-in memory? Built-in memory is unpredictable. The manual feedback log allows intentional review and deliberate rule creation without bloating CLAUDE.md.
• Findings: Feedback log caught issues like over-commenting, bad dev server teardown, and MCP worktree errors.

ROUND 2: DISSECTION | 血肉解剖

• Workflow Pipeline Breakdown:
  • Scoping & Kickoff: High-level grilling sessions to solidify plans, test plans, and smoke tests before any code is written.
  • Implementation & Self-Verification: Claude does the coding, but must verify its work via MCPs (Model Context Protocol) and self-review against predefined critical rules.
  • Review Stages: Draft PRs for initial checks, followed by a ‘Ready PR’ stage executing E2E tests, PR templates, and architecture reviews.
• The Feedback Mechanism:
  • Instead of relying on auto-memory, the author uses a TSV file.
  • Columns: date, tag, mistake, correction.
  • The 3-strike rule: If a tag appears 3 times, it signals a pattern, triggering a process to encode a formal, durable rule (like a check-rules.sh grep).
• Curation over Automation: The developer manually reviews this TSV log every 1-2 weeks. Only truly repeated and verified issues get promoted to the project’s permanent instructions (avoiding CLAUDE.md bloat for minor issues like Tailwind pixel values).

ROUND 3: SOUL | 靈魂提取

The essence of this approach is treating AI interactions not as one-off conversations, but as a system requiring intentional governance. By formalizing a feedback loop with a strict schema (the TSV log) and a threshold for rule creation (the 3-strike rule), the developer avoids the chaos of unpredictable AI memory and instead engineers a deterministic, self-improving development factory.

DEEP READ | 精讀指引

• “Because implementation is extremely strong everything feels weak in comparison.”
• “Durable corrections only - one per line, tab-separated… date -> tag -> mistake -> correction.”
• “A tag that reaches 3 lines is a pattern; /learn then proposes the smallest durable rule for it.”
• Difference from built-in memory: “memory will pick something up on its own right, which I don’t always like. I want a log of the feedback, so I get to review it.”
• Guarding against bloat: “For the things I don’t want bloating the CLAUDE.md, I add to critical paths which is picked up in self-review and pr-ready.”