Consolidate skills onto skills/ as the single source of truth

[groksrc]

Background

PR #859 vendors the Claude Code plugin, shared skills, Hermes, and OpenClaw into the monorepo. Now that everything is co-located, the two overlapping skill collections can be deduplicated. This split arrived organically and should be cleaned up — it is not an intentional design boundary.

Today:

• skills/ — 10 framework-agnostic memory-* skills (canonical source)
• plugins/claude-code/skills/ — 6 hand-maintained skills, several of which duplicate the shared ones

The tell: OpenClaw already treats skills/ as the single source (bun run fetch-skills bundles them). The Claude Code plugin is the outlier — it reinvents its own parallel set. Aligning the plugin to the same "bundle shared skills + add host-specific glue" pattern is the consistent architecture.

Target architecture

One canonical set of memory-* skills in skills/. Both hosts consume it identically:

• OpenClaw: bundles via fetch-skills (already done)
• Claude Code plugin: bundles the same shared skills, and layers on only the host-specific glue (the hooks)

Key reframe: most "plugin-specific" skills are actually shared content + a hook trigger. Separate the skill (agent-agnostic logic) from the glue (the hook), and almost everything collapses into skills/.

Disposition checklist

• research → retire, use shared memory-research (direct duplicate)
• edit-note → retire, fold into memory-notes (or a shared memory-edit if the edit workflow warrants standing alone)
• knowledge-organize → retire in favor of memory-defrag + memory-lifecycle — ⚠️ first verify they cover orphan-detection / relation-suggestion; migrate any unique content
• placement → move folder-decision logic to a shared memory-placement; keep the PreToolUse hook in the plugin (only the trigger is CC-specific)
• knowledge-capture → move to shared memory-capture; keep the JSONL session-ID detection as plugin glue (only that part is CC-bound)
• continue-conversation → decide: retire in favor of the MCP continue_conversation prompt + memory-reflect, or promote to a shared memory-continue

Naming

• Uniform memory-* namespace everywhere. The prefix earns its keep: shared skills install into flat dirs (e.g. ~/.claude/skills/) where memory-* avoids collisions with other vendors' skills — true regardless of host.
• Accept the minor cosmetic stutter in plugin slash commands (/basic-memory:memory-research); not worth diverging names to avoid.

End state

• A single memory-* skill set in skills/
• Claude Code plugin = hooks + a bundle of shared skills (mirrors OpenClaw)
• A single "Adding a New Skill" checklist that keeps dirs ↔ README ↔ lock/bundle in sync (see chore(installer): consolidate agent packages #859 review re: stale skills-lock.json)

Follow-up to #859 (not a blocker — co-location is the precondition for this dedup).

[claude]

Claude finished @groksrc's task in 2m 6s —— View job

---

Issue Triage

• Read and understand the issue
• Explore repository structure
• Classify and assess the issue
• Apply appropriate labels
• Provide triage summary

---

Analysis

Type: Enhancement (architectural refactoring / technical debt)

This is a code organization improvement that deduplicates overlapping skill collections following PR #859's monorepo co-location. It's not a bug and doesn't affect user-facing functionality.

Priority: Low

The issue itself notes this is a follow-up to #859 and explicitly "not a blocker." It's clean-up work that improves maintainability and architectural consistency — valuable but non-urgent.

Component: Plugin/skills architecture

Touches skills/ (shared memory-* skills), plugins/claude-code/skills/ (6 plugin-specific skills), skills-lock.json, and any bundle scripts like fetch-skills.

Complexity: Complex

Six individual skill dispositions each require distinct analysis:

• 2 retirements (verify replacements cover all use cases first)
• 1 fold-in (needs content migration)
• 2 promotions to shared + plugin glue split (architectural pattern change)
• 1 open decision (continue-conversation: retire vs promote)

The PR description is well-structured with a clear disposition checklist, so the work is well-scoped even though the execution is non-trivial.

---

Observations

• skills-lock.json is already present in the repo root — will need updating as skills are bundled
• The "bundle shared skills + add host-specific glue" pattern is sound: keeps agent-agnostic logic separate from hook triggers
• The naming consistency goal (memory-* everywhere) is worth enforcing to avoid flat-dir collisions in ~/.claude/skills/
• The continue-conversation disposition is the only open question — worth resolving early to avoid rework

Label applied: enhancement

[phernandez]

PR #865 (the v0.4 Claude Code plugin redesign, stacked on #859) already resolves the core of this — the plugin no longer reinvents a parallel skill set — though it took a different path than the checklist on a couple of items. Mapping it so this issue can be re-scoped:

Done in #865 (the 6 duplicates are deleted):

Skill	Checklist disposition	#865
research	retire → memory-research	✅ deleted
edit-note	retire → memory-notes	✅ deleted
continue-conversation	retire/promote	✅ deleted — replaced by the SessionStart brief + recall
knowledge-organize	retire → memory-defrag+memory-lifecycle	✅ deleted (see ⚠️ below)
placement	shared memory-placement + keep PreToolUse hook	🔀 diverged — dropped the skill and the PreToolUse hook; placement is now an output-style reflex + placementConventions in settings (the PreToolUse hook was intentionally removed — too noisy)
knowledge-capture	shared memory-capture + JSONL glue	🔀 diverged — replaced by the PreCompact checkpoint hook + /basic-memory:remember + the capture reflex in the output style

So placement and knowledge-capture didn't become shared skills — they became hooks + reflexes, which is the bridge architecture this plugin is built around. Worth deciding whether you still want shared memory-placement / memory-capture skills in addition (for non-CC hosts), or whether the hook/reflex form supersedes them.

Also in #865 — the "plugin bundles shared skills" end-state, done without repo duplication: /basic-memory:setup now installs the canonical memory-* set via npx skills add basicmachines-co/basic-memory --path skills (same source OpenClaw bundles). The plugin stays hooks + CC-specific skills and pulls the shared toolkit on demand — no vendored copies in the repo.

Naming: keeping the 4 CC-specific skills un-prefixed (/basic-memory:setup, :remember, :share, :status) since the plugin namespace already carries basic-memory. The memory-* prefix still applies to the shared set (flat-dir install). So no rename needed on the plugin side.

Belongs in a separate skills-package issue (not the plugin):

• ⚠️ Verify + migrate before retiring — feat(plugins): Claude Code plugin v0.4 — the memory bridge #865 deleted knowledge-organize outright; someone should confirm memory-defrag/memory-lifecycle actually cover its orphan-detection / relation-suggestion, and migrate anything unique. Same spirit for any unique placement/knowledge-capture content.
• The skills-lock.json staleness + the single "Adding a New Skill" dirs↔README↔lock checklist.

Net: the dedup is effectively done in #865; what remains is (a) the design call on whether memory-placement/memory-capture should exist as shared skills given the hook/reflex replacements, and (b) the skills-package housekeeping (migration verification + lockfile/checklist), which is cleaner as its own issue.

[phernandez]

goal is to remove plugins/claude-code/skills/