[lukemaj]

What: The README, overview, AGENTS.md, update-mechanism page and example Skill state the repo's current scope and point at Issues, and the loader Skill is now a routing table whose paths resolve from its own base directory.
Why: The entry points described a karpathy-wiki-only repo, required reading a superseded handoff, and shipped a loader whose bare docs/... paths did not resolve from an installed Skill (the relative-path failure behind the toolboxmd/agentsmd#164 pilot escape).
So what: The planner runs Luna max review on the exact head; the user merges. A behavioral trigger test of the loader stays with #17.

Closes #14
Refs #7

Changes

• skills/building-agentskills/SKILL.md: rewritten as a routing table, 24 rows, one before-clause and one page per row, one hop. The top says once that every path is relative to the Skill's base directory, and to read the same path from raw.githubusercontent.com/toolboxmd/building-agentskills/main/ when the page is absent (a copied Skill). It asks for the read as its own step, before any command or edit (the #182 lesson in triggers.md). The description names the actions (411 characters). The Iron Law now matches the AGENTS.md evidence contract. The summary map, the three-question summary and the 500-line section are removed; their pages are rows. This makes 08-packaging-as-plugin.md ("follows convention 3") true without editing it.
• README.md, docs/00-overview.md: scope is authoring, routing and trigger benchmarks, with karpathy-wiki and AgentsMD as evidence sources. "Claude Search Optimization (CSO)" becomes "Skill Discovery Optimization (earlier called Claude Search Optimization)", per triggers.md. The README map now describes Codex, Gemini CLI and others.md (Grok Build) as Correct the harness facts for Claude Code, Codex, Grok Build, OpenCode and Gemini CLI #8 left them, adds a trigger-benchmark audience line, and drops "~30-line". Source nicknames still used (LANDSCAPE, REVIEWER, LESSONS) are already defined in the overview's Sources section.
• AGENTS.md: "Start here" drops the handoff and names Verify the loader Skill is discoverable and triggers on authoring and audit requests #17 to Gate version anchors and CHANGELOG entries in npm test #22; "Current local drafts" names TODO.md and the drift draft as user-owned untracked files whose entries live in those Issues; new Proof section (npm ci && npm test, run by CI); wiki/ is maintained by karpathy-wiki and skipped by the doc checks.
• docs/12-update-mechanism.md: three case studies, Grok Build coverage, the trigger-benchmark method; "planning backlog" now links GitHub Issues; last landscape audit dated 2026-04, platform facts refreshed 2026-09-29 by Correct the harness facts for Claude Code, Codex, Grok Build, OpenCode and Gemini CLI #8.
• examples/minimal-skill/SKILL.md: the triggers link is a GitHub URL; the Iron Law text says a trigger is necessary, not sufficient, citing the 0/40 to 20/20 result.
• Version 0.1.0 to 0.2.0 in .claude-plugin/plugin.json, package.json and the root entries of package-lock.json (regenerated with npm install --package-lock-only; a lock left at 0.1.0 would contradict package.json). The plugin description states the new scope. public/llms.txt regenerated.

Elon record

• Requirements and who asked: user, 2026-09-29 (Clean up the repository and apply the routing and benchmark learnings #7, Make the README, overview, AGENTS.md and loader Skill state the current scope and route by action #14): entry points state the current scope, point at Issues, and the loader routes by action with paths that resolve from its base directory, applying the routing lessons (Correct triggers.md and add the trigger-benchmark method #5, Case study: the AgentsMD routing benchmarks (toolboxmd/agentsmd#164, #174, #182) #6) to the loader itself.
• Deleted: the loader's doc map, three-question summary and 500-line section (the pages are rows); the required reading of the superseded handoff; a behavioral trigger benchmark of the loader (Verify the loader Skill is discoverable and triggers on authoring and audit requests #17).
• Bottleneck: the first files a reader or agent opens sent them to stale plans and unresolvable paths.
• Checked myself: every loader path resolves (below); breaking one loader path makes tests/check-docs-links.test.sh fail, so the link check covers the loader; description length 411 characters; npm ci && npm test under Node 22.23.1.

Proof

• npm ci && npm test at 1ad36ae (rebased onto 20eed52, which merged docs: routing and harness anti-patterns and reword-first evidence (#13) #27), Node v22.23.1: all six test scripts pass. Same result at the first head 0f7590e (build-llms, check-docs-links, check-llms-coverage, no absolute paths, bare-paths gate, provider-neutral wiring).
• Loader paths, from skills/building-agentskills/: for p in $(grep -oE '\(\.\./\.\./[^)]+\)' SKILL.md | tr -d '()'); do test -f "$p" && echo "ok $p" || echo "MISSING $p"; done prints 24 ok lines, 0 MISSING.
• Negative check: replacing ../../docs/09-evolution.md with ../../docs/09-nope.md makes bash tests/check-docs-links.test.sh fail with skills/building-agentskills/SKILL.md ⎿ ../../docs/09-nope.md; restored afterwards.

Not changed / follow-ups

• Add the routing and harness failure modes to the anti-pattern catalog and the mechanism page #13 merged (docs: routing and harness anti-patterns and reword-first evidence (#13) #27); the README map entry for 10-anti-patterns.md now names its routing and harness entries. State the measured activation behavior in the quickstart, mental model, three questions and token economics #12 (docs: activation and budget pages match measured behavior (#12) #28) is still in review: the loader and README link its pages (01 to 04) at page level only, with action triggers that do not depend on the content it changes. Follow-up: once docs: activation and budget pages match measured behavior (#12) #28 merges, check that the README descriptions of those pages still match.
• The hand-kept intro of public/llms.txt above the generated block still says "Layered authoring patterns"; the ticket treats the file as generated only, so it is left as is.
• No versionctl policy exists in this repo; the release process is Gate version anchors and CHANGELOG entries in npm test #22.

🤖 Generated with Claude Code

[lukemaj]

Independent review: APPROVE

Reviewed exact head 0f7590ec42fac28165df7c72adef72d0b1e1e94c against Issue #14. No blocking findings.

• All Issue acceptance criteria and the declared file scope are met. The root package-lock.json version entries follow the declared package version bump, which the PR body explains.
• The loader names each action in a before-clause, routes each row to one page, asks for the page to be read as its own step before action, and resolves all 24 paths relative to its base directory with a GitHub fallback. The 411-character description names the covered actions, and its Iron Law and evidence paragraph match the AGENTS.md evidence contract.
• Checked changed text for em dashes and /Users/... paths; none found.
• From a fresh worktree at the reviewed SHA, Node v22.23.1: npm ci && npm test passed. All 24 loader paths passed test -f from skills/building-agentskills/.

Reviewed PR: #29

[lukemaj]

Re-reviewed rebased head 1ad36aed252acddb20e7ce9f626dfaf90ab65789: APPROVE. The README map wording matches docs/10-anti-patterns.md entries on routing failures and benchmark scoring errors. In a fresh worktree, Node v22.23.1 npm ci && npm test passed; all 24 loader paths resolve from skills/building-agentskills/.