Live Enter an objective; the design run executes server-side against live databases and Claude, then streams its result and the annotated neochromosome back here.

Section 03 — how this was built

Built with
Claude Code

NeoFlex Designer — an eight-module pipeline that turns a biological objective into an orderable, in-silico neochromosome — was designed and written end to end with Claude Code. Every figure below is read from the project's own usage tracker, session transcripts, and git history, and checked for internal consistency, not estimated.

Token usage

usage/summary.json

5.43Btotal tokensinput + cache + output across the whole build

23.67Mgenerated · fresh model outputthe tokens Claude actually wrote — 0.4% of the total

Code written

git log · numstat

161,883Python lines addedcumulative additions across 350 commits, initial import excluded

105,085Python in final modules452 module files, present-tree wc -l

Python lines added per day · cumulative to 161,883

Commit history

git log · 351 commits

Commits per day · cumulative to 351

Work by area · commit-message prefix

Tools & orchestration

4,488 tool calls

Tool calls · top 12 of 46

Subagent runs · 1,046

Workflows · 83

How it was built

the working pattern

Ideate AutoSci literature wiki

A queryable literature wiki seeded the design

Ideation ran on a self-hosted LLM literature wiki (AutoSci, a vendored MIT research-agent toolkit). Claude cross-linked the neochromosome, CAZyme and metabolic-engineering literature into a searchable wiki that seeded the DesignSpec and M1's evidence-backed discovery, so parts trace to real accessions and DOIs, not model recall.

Scale Waves of Opus subagents

1,046 Opus subagents, in rolling waves

Work was fanned out in rolling waves of roughly eight concurrent Opus subagents. Counted from the agent transcripts on disk, 1,046 subagent runs executed — 922 spawned inside the orchestrated workflows and 124 launched directly; the tracker's headline of 105 counted only the directly-launched calls, not the agents each workflow fans out. 99.7% of all 11,084 turns ran on Claude Opus 4.8, across explicit wave13-e2e / wave17-e2e checkpoints.

Track Kanban board

A living board, auto-updated on every commit

A module status board under kanban/ tracked working / partial / missing features and test counts per module, regenerated by a git post-commit hook. It is the most-touched artifact in the repo: 73 of 351 commits carry the kanban: prefix.

Verify Review & goal-driven gates

Changes were gated to a go / no-go

Dedicated review, audit and go-review passes (14 commits) plus 83 workflows fed a reviews/ tree of findings, catching fabricated citations before an order package. Modules talk through 12 versioned contract: schemas validated by a deterministic harness; runs looped until the fail-closed gate cleared.

Module status board

05_project_code/kanban

The living kanban board that tracked the eight-module build — working / partial / missing features and test counts per module, regenerated by a git post-commit hook on every commit. It is the most-touched artifact in the repo (73 of 351 commits carry the kanban: prefix). Self-contained page (board.html), so it loads on demand.

Usage counts read from 04_Claude_usage_tracking/usage/summary.json (snapshot 2026-07-13T13:01Z); commit and line-of-code figures from git log over 351 commits (2026-07-14). The 1,046 subagent runs are counted directly from the workflow agent transcripts under ~/.claude/projects/*21-Hackathon*/…/subagents/ (1,046 agent records: 922 spawned inside workflows + 124 launched directly), because the tracker's summary logs only the 105 top-level Task calls, not the agents each workflow fans out. Raw copies ship beside this page as assets/usage_summary.json; the module status board ships as board.html.