Dev builds AI coding tool using itself across 88 sessions and 201 commits
The author of Wrought — a structured engineering process tool for AI coding assistants — spent 87 days dogfooding their own product across 88 Claude Code sessions and 201 commits, documenting hard-won lessons about building software with AI.
Score breakdown
Practitioners building with AI coding assistants can adopt the Findings Tracker pattern — structured markdown lifecycle files with dependency maps and artifact links — to maintain continuity across sessions and avoid rediscovering prior work from scratch.
- 0188 Claude Code sessions and 201 commits over 87 calendar days (Jan 26 to Apr 23) to build Wrought
- 02The project has ~2,300 lines of Python source but over 193,000 lines of total insertions — an 84-to-1 ratio dominated by documentation and process artifacts
- 03Test code (~3,100 lines) exceeds production source code (~2,300 lines)
The post by Wrought's author chronicles 87 days of building Wrought — a structured engineering runbook for AI coding assistants — using the product itself for every session. The raw metrics are striking: 88 Claude Code sessions, 201 commits, 47 findings trackers, 60 design documents, 55 blueprints, 33 research reports, 23 parallel code reviews, 75 implementation prompts, ~2,300 lines of Python source, ~3,100 lines of test code, and over 193,000 lines of total insertions. The 84-to-1 ratio of insertions to production source code is the central data point: the project is mostly documentation, process records, and methodology artifacts rather than code.
This directly addresses AI coding assistants' core limitation: they forget everything between sessions.
The most valuable pattern to emerge was the Findings Tracker — a markdown file that tracks work through a structured lifecycle (`Open -> Investigating/Designing -> Blueprint Ready -> Planned -> Implementing -> Resolved -> Verified`), complete with dependency maps, resolution tasks with checkboxes, lifecycle timestamps, and links to every artifact produced. This directly addresses AI coding assistants' core limitation: they forget everything between sessions. A concrete example given is the "Context Compaction Resilience" tracker, which managed a five-sub-finding, four-session investigation into Claude Code's auto-compaction destroying in-flight state, ultimately producing a five-layer defense architecture. Across 47 such trackers, the project maintained continuity that would otherwise have been lost at each session boundary.
The post also challenges the intuition that faster code generation reduces the need for design. The author found the opposite: when code generation is nearly free, skipping analysis produces code that works but is architecturally questionable, solves the wrong problem, or makes future features harder to build. The pipeline Wrought now enforces is `/research -> /design -> /blueprint -> /wrought-implement -> /forge-review`, ensuring every feature is researched, analyzed with a structured tradeoff matrix, and fully specified before a single line of implementation begins.
Key facts
- 0188 Claude Code sessions and 201 commits over 87 calendar days (Jan 26 to Apr 23) to build Wrought
- 02The project has ~2,300 lines of Python source but over 193,000 lines of total insertions — an 84-to-1 ratio dominated by documentation and process artifacts
- 03Test code (~3,100 lines) exceeds production source code (~2,300 lines)
- 0447 'Findings Trackers' — structured markdown files tracking work through a defined lifecycle — serve as the project's institutional memory across sessions