claude-memory: aihero course lane 9: steering-surface validations (navigation pointers, harness claims, auto-memory, /init-then-prune eval)

[kyle-sexton]

Context

Steering-section extension of the AI Hero course vetting effort — contract in docs/topics/pocock-course-lanes/PLAN.md (branch claude/plan-mode-discussion-55kszx); lanes 1–6 are #2899–#2904; sibling steering lanes are #2909 and #2910. This lane sweeps the steering lessons whose territory we believe is already covered, validates that belief instead of assuming it, and verifies the section's harness-behavior claims per the contract's claim ladder.

Expected-parity surfaces to validate, not assume: claude-memory:audit (memory-layer hygiene), claude-memory:stateless (auto-memory), docs-hygiene:rename-references (stale pointers), docs-hygiene pruning machinery.

Proposed work

• Open with /planning:interview me per contract; lane discusses and decides — implementation leaves as filed work items
• Navigation pointers (lesson 6): check claude-memory:audit criteria for coverage of: navigation sections in CLAUDE.md, stale-pointer risk ("a stale highway is worse than no highway" — pointer targets that moved), nested/subdirectory CLAUDE.md loading (the course doesn't mention it; our guidance should), and @-mention as the one-turn-scoped equivalent. Patch criteria if absent
• Design-smell caveat (user-raised): a navigation pointer that exists because changes must be mirrored across distant folders can mask low cohesion ("things that change together live together") — remediation guidance should consider restructuring before adding a pointer. Decide where this caveat lives
• Auto-memory (lesson 9): validate claude-memory:stateless + audit fully cover the lesson (expect no work; record the parity rows)
• /init-then-prune eval fixture (lesson 8, user-suggested): evaluate generating a bloated memory file with /init, running claude-memory:audit (or the lane-7 authoring skill) against it, and grading against expected findings — as an eval fixture / benchmark flow (evals:design)
• Harness-claims bundle (claim ladder tier i — verify against current official docs before repeating): auto-memory file location (per-project state dir keyed by cwd, outside the repo), /memory command behavior, MEMORY.md index shape, disable-model-invocation listing behavior, nested CLAUDE.md loading semantics, ~/.agents/skills vs ~/.claude/skills path equivalence claim, agentskills.io open-standard claim, /context Memory Files accounting
• Pruning terminology note: "sediment" / "cache" / "no-op" adoption is lane 6 (docs(upstream): aihero course lane 6: shared vocabulary adoption and aihero-course provenance doc #2904) term-adoption territory — feed candidates there, do not decide here

Acceptance criteria

• Adopt/reject-with-reason/track-on-event decision rows in docs/upstream/aihero-course.md for lessons 6 (Navigation Pointers), 8 (Trying Out Pruning), 9 (Automatic Memory), plus leftover claims of lessons 1–2 not owned by playbooks/docs-hygiene: aihero course lane 7: authoring-time steering doctrine (writing-for-agents re-decision) #2909
• claude-memory:audit criteria gaps decided (patched or explicitly declined with reason)
• Design-smell caveat homed or rejected with reason
• Harness-claims verdicts recorded with corroboration labels (same table shape as the contract's existing harness-claims section)
• Eval-fixture verdict recorded; work item filed if adopted
• Term candidates handed to lane 6

References

• docs/topics/pocock-course-lanes/PLAN.md (claim ladder; harness-claims verdict table shape)
• plugins/claude-memory/skills/audit/SKILL.md + reference/criteria.md
• plugins/claude-memory/skills/stateless/SKILL.md
• docs/upstream/mattpocock-skills.md — decomposition table

Metadata

Field	Value
Category	unspecified
Area	claude-memory, docs-hygiene, evals, docs/upstream
Ecosystem	unspecified

[kyle-sexton]

Claimed for lane 9 of the steering-lane chain (session continuation of the chain that closed #2909 and #2910; user-directed in-order execution, auto-confirmed — no interactive selection prompt). Claim via MCP assignee (verify-assignee-then-proceed; no gh CLI in this container, seam lease unavailable — same degradation recorded on #2910). Work proceeds on claude/pocock-steering-course-00zkvd; lane discusses and decides per the contract — implementation leaves as filed work items. Note: decision rows will land in docs/topics/pocock-course-lanes/STEERING.md per the chain convention (not docs/upstream/aihero-course.md, which stays lane 6's deliverable, #2904) — deviation from the first acceptance criterion's stated path, to be recorded at close.

---

Generated by Claude Code

[kyle-sexton]

Lane 9 closed. Interview-first per contract (rounds 1–2, all decisions user-confirmed); decision rows and status: docs/topics/pocock-course-lanes/STEERING.md (lane 9), commits fabce556 + 0326845c on claude/pocock-steering-course-00zkvd.

Against the acceptance criteria:

• Decision rows (lessons 6, 8, 9 + L1–2 leftovers) — recorded in STEERING.md per the chain convention (deviation from this issue's stated docs/upstream/aihero-course.md target, which stays lane 6's deliverable, docs(upstream): aihero course lane 6: shared vocabulary adoption and aihero-course provenance doc #2904 — noted at claim time). Lesson 6: C5 carve-out + C7 stale-highway note adopted; nested-CLAUDE.md placement row and @-mention one-turn-scope distinction adopted as extensions. Lesson 8: eval fixture adopted. Lesson 9: PARITY, no work. L1–2 leftovers: verdict-backed rows only, including an explicit do-not-repeat row on the refuted ~/.agents/skills path.
• claude-memory:audit criteria gaps decided — patched via filed item claude-memory: patch audit criteria for navigation pointers — C5 carve-out, C7 stale-highway note, C3 nested CLAUDE.md + @-mention rows, design-smell remediation #2987 (lane does not implement).
• Design-smell caveat — homed in claude-memory: patch audit criteria for navigation pointers — C5 carve-out, C7 stale-highway note, C3 nested CLAUDE.md + @-mention rows, design-smell remediation #2987's remediation guidance; authoring skill pointed at it via coordination comment on docs-hygiene: implement write-for-agents — authoring-time skill for agent-consumed markdown #2962.
• Harness-claims verdicts — 7-claim table in STEERING.md, contract table shape, graded against Claude Code v2.1.233 with corroboration labels (verifier-corrected: claim 4 single-pool docs-only with cure-when-convenient note; ~/.agents/skills refuted two-pool). Lane-8's disable-model-invocation listing verification reused, not repeated.
• Eval-fixture verdict — ADOPT; filed as claude-memory: /init-then-prune eval fixture — static bloated CLAUDE.md graded against expected audit findings #2989 (static bloated-CLAUDE.md fixture against the audit eval suite, determinism over live /init).
• Term candidates → lane 6 — "highway / stale highway" increment posted on docs(upstream): aihero course lane 6: shared vocabulary adoption and aihero-course provenance doc #2904.

Steering lanes 7–9 are all closed; the chain proceeds to PR + merge.

---

Generated by Claude Code