- introduce active skill echoing to prevent context window forgetting during execution
- enforce strict state machine transitions back to planning mode after execution
- replace AI apologies and empty promises with mandatory post-mortem rule updates
- emphasize quality over speed as a core behavioral principle
Implement a mandatory State Machine model (PLAN and IMPLEMENTATION states)
in root instructions to prevent the AI from generating code prematurely.
The rules require mechanical reloading of skill files during execution
to combat context-window memory loss.
Also update Mermaid diagram instructions to explicitly ban <br> tags for
newlines and prohibit raw angle brackets (<, >) in nodes, enforcing
real newlines and HTML entities (<, >) to fix rendering issues.
-root relied on AI already having read the instructions — useless for
session start. Replaced with a direct read instruction at the top of
README that works regardless of prior context.
config.yaml: new repo-root config file with docs_folders list
(docs, documentation, doc). apply.sh reads this list and picks the
first existing folder per project instead of hardcoding docs/.
Instructions:
- core-principles: add No Vibe Coding and No Touching .ai/ sections
- ai-root-instructions: add mandatory instructions block — rules stay
active for the whole session, not just at start; AI must stop and
announce if instructions were not loaded
- project-context, docs: updated to list all docs folder alternatives
and reference config.yaml as the source of truth
FR-5.0 added to apply-requirements.md. README step 4 updated.
New skill files:
- clean-code.instructions.md — naming, functions, classes, error handling,
formatting, DRY/KISS/SOLID, Helm YAML conventions
- clean-architecture.instructions.md — dependency rule, layers, boundaries,
SOLID foundation, Helm as outermost layer
- helm.instructions.md — resource ownership, values hygiene, required vs
defaults, umbrella chart pattern, two-file values layering, KISS principle,
hook ordering, config files pattern, dependency caching, template testing
Register all three in ai-root-instructions.md skills list and routing table.
Remove .ai from .gitignore — .ai/ is the product in this repo and must be
tracked; the apply.sh skip-by-marker mechanism prevents changes to other repos.
Remove QUICK-REFERENCE.md from .ai/ — it contained kube-log project-specific
content (hardcoded paths, OpenShift/ArgoCD config) that was leaking into all
projects via the symlink.
Update docs to match the actual apply.sh implementation:
- apply.md: rewrite to reflect curl-only invocation, bootstrap/local phases,
factory-reset marker behaviour; remove .ai-instructions.conf and interactive
checklist references
- architecture.md: remove scripts/ subdirectory and .ai-instructions.conf that
were never implemented
- apply-usecases.md: remove INSTALL subgraph (rm -rf + cp -r was never done),
fix symlink targets to SCRIPT_DIR/.ai
- apply-requirements.md: NFR-1 corrected (TTY-gated prompt exists), NFR-4
updated to <150 lines, .ai-instructions.conf removed from out-of-scope list
