claude-config: add model-upgrade unhobbling audit — sweep instruction surfaces for constraints newer models no longer need

[kyle-sexton]

Purpose

Upgrade-triggered (or on-demand) audit that sweeps the local Claude Code instruction surfaces and proposes removals/rewrites of deterministic constraints that hobble newer models. Source: Anthropic's Fable 5 prompting guidance ("re-evaluate which instructions, tools, and guardrails are still needed"; "Skills developed for prior models are often too prescriptive… and can degrade output quality") + Thariq Shihipar's 80% system-prompt-cut account.

Audit surfaces

Same doc-class, same checks:

• user + project CLAUDE.md
• skill SKILL.md bodies (+ context files)
• agent definitions
• .claude/rules/**/*.md
• prompt-type hooks and output styles

Check catalog (seeded, sourced)

11 sourced seeds captured in knowledge-corpus .work/youtube-watch/how-i-plan-build-and-run-loops-with-clau-aVO6E181cNU/research/findings/context-trimming.md. Headline checks:

• Bare prohibitions ("never/do not X" without why) → prohibition-plus-rationale rewrite, not deletion
• Over-prescriptive step lists → cull to intent + constraints
• Constraining example blocks → keep 3–5 format-steering examples (docs still recommend those; no blanket ban)
• Every instruction against the pruning bar: "Would removing this cause Claude to make mistakes?" (code.claude.com/docs/en/best-practices)
• Stale model-era workarounds (instructions compensating for prior-model weaknesses)

Design decisions (locked in interview session claude-loop-practices, 2026-07-21)

• Home: claude-config audit family.
• Output: findings report + proposed diffs, human-gated — never auto-applies.
• Cross-repo routing: findings inside standards-managed materializations route upstream to melodic-software/standards per sync-manifest ownership; never edited in place.
• Composes with (distinct intents, pointers only): skill-quality:check (structural lint), docs-hygiene:compress (token brevity), claude-config:audit (config hygiene).

Key sources

• https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5
• https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices
• https://code.claude.com/docs/en/best-practices
• 80% cut (secondary, no engineering post exists): https://the-decoder.com/anthropic-says-it-cut-80-percent-of-claude-codes-system-prompt-because-fable-5-models-want-a-smaller-system-prompt/

Pipeline

Brief (topic slice) → /discovery:explore → /planning:plan → implement.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y4RMR3G1rHAdNjxTqenVJL

[kyle-sexton]

Interview follow-up (2026-07-21) — naming + execution shape:

• Naming: "unhobbling" is source vocabulary (Thariq's talk part 1), too obscure for a skill name. Run /naming:name-it-better at pickup; candidates to seed: instruction-audit, trim-instructions, constraint-audit.
• Execution shape — per-surface lanes, not one monolithic pass: sub-agent lane per surface (CLAUDE.md, rules, skills, agents, hooks) sharing one check catalog; large lanes fan out further (one agent per skill). Lanes run/loop independently so each gets real depth.
• Scope bound: audit only locally-owned surfaces (user CLAUDE.md, ~/.claude/rules, project .claude/). Marketplace plugin content is upstream-owned — findings there route to claude-code-plugins issues, never local cache edits.
• Verify pass: every proposed removal gets an adversarial "would removing this cause mistakes?" check before it reaches the human-gated diff — removal is cheap to propose, expensive when wrong.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y4RMR3G1rHAdNjxTqenVJL

[kyle-sexton]

This was generated by AI during triage.

Triage: briefed (delegable) → status: ready. Feature request; design locked by the operator's interview session (claude-loop-practices, 2026-07-21). Type Feature. Priority medium (claude-config outside the week's priority-plugin set). Not a dup. No open prerequisite.

Behavioral contract. Add a claude-config audit-family skill: an upgrade-triggered or on-demand audit that sweeps the local Claude Code instruction surfaces — user + project CLAUDE.md, skill SKILL.md bodies + context files, agent definitions, .claude/rules, prompt-type hooks and output styles — and proposes removals/rewrites of deterministic constraints that hobble newer models. Output = a findings report + proposed diffs, human-gated, never auto-applied.

Locked design (operator interview — do not re-open):

• Check catalog seeded from the sourced findings: bare prohibitions ("never/do not X" without why) → prohibition-plus-rationale rewrite (not deletion); over-prescriptive step lists → cull to intent + constraints; keep 3–5 format-steering examples (not a blanket ban); the "would removing this cause Claude to make mistakes?" pruning bar; stale model-era workarounds.
• Cross-repo routing: findings inside standards-managed materializations route upstream to melodic-software/standards per sync-manifest ownership — never edited in place.
• Composes (pointers only, distinct intents): skill-quality:check (structural lint), docs-hygiene:compress (token brevity), claude-config:audit (config hygiene).

Pipeline: standard Brief → discovery/explore → planning/plan → implement; the planning stage gates implementation specifics.

[kyle-sexton]

[kyle-sexton]