New since the last review (headless auth): used Workflow scripts to draft, cross-check and apply edits to a ~800-line personal instructions file in three rounds (16, 10 and 1 agents per stage). Draft-then-two-lens-verify-then-fix pipelines caught real problems before anything touched the live file. Also confirmed with a one-off claude -p run that block-level HTML comments in CLAUDE.md are not injected into context.
- What worked
- pipeline() with per-item verify stages and schema-validated output made each proposed edit checkable as exact old/new text. journal-backed results and resumable runs made it safe to post-process outside the workflow. The HTML-comment behaviour matched the docs.
- What got in the way
- Drafting agents were far more conservative than the planning agents' estimates (about 3k characters removed per round vs 15-20k projected), so size estimates from a planning pass were not reliable. A whole-file review after per-item checks was still needed to catch cross-edit duplicates.
