Graft v6.2: Using Graft in Coding Workflows
Graft v6.2 refines the positioning and adds graft import. The key insight: Graft doesn't need to orchestrate everything — it handles the natural-language sub-steps within larger workflows.
The Hybrid Pattern
Coding tasks mix NL steps (planning, review, analysis) with code execution (file writes, tests, builds). Graft handles the NL parts:
Plan → Code (direct) → graft run review.gft → Fix (direct) → Done
↑
Graft handles this part
The bottleneck in coding tasks isn't harness engineering — it's filesystem side effects and iterative test loops. Graft's 77% token savings apply where agents exchange structured JSON, not where they write files.
So instead of trying to orchestrate the entire coding workflow, use Graft for the steps where it actually helps: review pipelines, analysis, planning, ideation. Run code execution directly through Claude Code.
graft import
New command: reverse-compile existing .claude/ structures into .gft files.
graft import ./my-project -o pipeline.gft
Reads agent definitions, hooks, settings, and orchestration plans from .claude/ and generates valid .gft source. Useful for converting manually-configured pipelines to Graft.
Positioning Update
All docs now reflect the refined scope:
- Full NL pipelines — code review, ideation, content generation, data analysis, debate architectures
- NL sub-steps in coding workflows — review, analysis, planning pipelines that run between code execution steps
- Not for — direct code execution orchestration (use Claude Code directly for that)
Numbers
- 1,722 tests
graft importwith round-trip verification (import → check passes)- Documentation updated across README, guide, index, and CLAUDE.md