QRHHomelab Quick Reference Handbook
PeriodicP5

Re-run prompt-audit on the dev plugin

When
After a model change: the model behind any dev-plugin role changes. Run it before reading the first slices run on the new model.
AIWorkflow Edit on GitHub

A new Opus or Sonnet release counts even with no edit at all: MODELS in run_loop.py / plan_loop.py uses the opus / sonnet aliases. Four agents pin a model in frontmatter: rebase-agent, test-agent and test-fixer on sonnet, refinement-writer on fable.

Steps #

  1. In AIWorkflow, run /claude-api prompt-audit (the skill's shared/prompt-audit.md procedure) over plugins/dev/agents, plugins/dev/skills, plugins/dev/docs and the dispatch prompts in run_loop.py / plan_loop.py.
  2. Re-read every finding at its file:line, and drop what doesn't hold.
  3. Apply the survivors as one plugin version, with a CHANGELOG-workflow.md entry.
  4. Add a row to the run log below.

Why #

Instructions drift relative to the newest model, and a frontier model executes old patches literally (Anthropic's platform-cost post, 2026-09-08).

Run log #

Date Models Findings Applied in
2026-09-09 Opus 5, main roles AIWorkflow docs/research/platform-cost-read-2026-09-09.md § 3.2 0.9.33
End of procedure