feedback_skill_discipline_2026_04_26

Rule: Invoke superpowers skills (especially using-superpowers / brainstorming / verification-before-completion) BEFORE acting, not as background knowledge. The skill content carries the discipline; memory + good intent do not. Knowing ≠ using.

Why: On Sunday 26 April 2026 Rich diagnosed verbatim: “this project has been moving very slowly, due to inadequate use of claude skills. in particular we need to user the brainstorming and using-superpowers and tdd/verification skills more.” Too many processes had been taking far too long. The pre-existing instructions — global CLAUDE.md §7 “Superpowers Plugin — MANDATORY” paragraph + the using-superpowers skill itself — were prose-only without concrete boundary tests, and they had not been enough. The dominant failure mode was the “I know what brainstorming is” red-flag flagged in using-superpowers itself: the discipline was being treated as background knowledge instead of explicit skill invocation, so the option-widening + verification + TDD checklists never actually loaded into context.

How to apply (boundary tests — if true → STOP and load):

• Session start + after any context compaction → re-invoke superpowers:using-superpowers. Skills silently drop post-compaction (6 April 2026 incident in global CLAUDE.md §7 still applies).
• About to generate >1 option / scorecard / plan / EnterPlanMode / “what should we do about X?” answer → invoke superpowers:brainstorming FIRST. If you find yourself producing options without prior brainstorming, STOP.
• Next sentence contains “done / pushed / passed / saved / locked / state file updated / in Downloads” → invoke superpowers:verification-before-completion and run the verification commands BEFORE writing the claim. This is the operational arm of Section 13 (No Lies).
• Schema, Catala rule, LinkML generator, CI gate, or any executable artefact change → superpowers:test-driven-development, failing test first — even for schema validators.
• Drift / count mismatch / scorecard-vs-state divergence / module-name disagreement → superpowers:systematic-debugging BEFORE proposing fixes.

Files updated Sunday 26 April 2026:

• ~/testatetech/docs-strategy/CLAUDE.md — new ”⚡ MANDATORY skill usage” section inserted at the top, ahead of the canonical-doc reading list. Contains boundary-test table for the three gating skills + secondary-skill list + anti-pattern catalogue + rule of thumb.
• ~/.claude/CLAUDE.md — Section 7 “Superpowers Plugin — MANDATORY” subsection rewritten with parallel structure (boundary-test table + domain-shaped skills + recurring anti-patterns + rule of thumb). The 6 April 2026 incident anchor preserved as load-bearing prose.

Specific anti-patterns that justified the strengthening (recurring in docs-strategy):

• Producing scorecards without invoking brainstorming to widen the option-set first → top option only emerges after Rich asks “dig deeper” (pattern recorded in feedback_automatic_deep_dive_when_options_feel_imperfect)
• Claiming “locked + committed + pushed + state file v{X.Y} updated” without verification-before-completion actually running git log --oneline origin/main -1 and confirming the lastmod field
• Resuming long sessions post-compaction without re-invoking using-superpowers — skills silently drop
• Treating “I know what brainstorming is” as substitute for Skill tool invocation
• Batching multiple questions to Rich in one turn (memory feedback_scorecards_one_at_a_time_optimal_sequence triple-elevated; skills reinforce, do not exempt, the single-question rule)

Boundary test for this memory itself: If you are about to produce strategic output in this repo and skill content for the relevant work-shape has NOT been loaded in this session — STOP and load it via the Skill tool or slash-command. One extra tool call < re-doing work that skipped a discipline. Slow progress = under-invoked skills.