Let your AI agent auto-upgrade your workflow weekly

One skill drives the full loop: pull the latest LoreJump-maintained best practices β†’ scan your project's harness state β†’ use your local Claude Code for DeepResearch β†’ preview and apply local changes β†’ generate a local report.

One run = 7-step closed loop

  1. 1
    Scan local project β€” reads CLAUDE.md / .claude/skills / .claude/hooks / .mcp.json / LoreJump score history, aggregates into a project fingerprint
  2. 2
    MCP pulls the latest practices β€” calls lorejump-mcp get_sota_pack for SOTA profile + evolution highlights + top-N upgrade candidates + a gap list tailored to your project
  3. 3
    DeepResearch suggestion blocks β€” emits WebFetch / Agent command templates per candidate; you decide whether Claude Code runs cross-verification
  4. 4
    Change preview β€” every candidate maps to a concrete diff / command; conservative mode previews only by default
  5. 5
    Apply (after you confirm) β€” selectively applies; rolls back any failed item; never touches src/ application code; respects local hook protections
  6. 6
    Generate local report β€” writes .lorejump/harness-cycle-YYYY-MM-DD.md (applied / skipped / needs-your-review lists)
  7. 7
    Receipt sent back β€” applied results and skip reasons go back via MCP submit_report (apply / session_close markdown receipts) to LoreJump, driving next week's practice library

Install (requires lorejump-mcp first)

Prerequisite: lorejump-mcp connected

Claude Code (1 step)

curl -fsSL https://lorejump.com/install/lorejump-harness.sh | bash

The script places SKILL.md at ~/.claude/skills/lorejump-harness/

Run

# Preview (conservative mode, default)
/lorejump-harness

# Apply
/lorejump-harness --apply

# Only what's been added since a given date
/lorejump-harness --since 2026-04-13

# Deep verification (extra WebFetch/Agent cross-check round)
/lorejump-harness --deep

Example output (not a real project)

πŸ”§ LoreJump Harness upgrade preview β€” 2026-04-20
━━━━━━━━━━━━━━━━━━━━━━━━━
5 candidates (sorted by priority_score)
── 1. [CC-2026-W17-skills-token-budget] Skill 2K-token cap best practice (priority=89) ──
Applies to: context_management + multi_agent (your D2=4 β†’ target D2=5)
Source: anthropic.com/skills-authoring-guide
Change preview: edit .claude/skills/foo/SKILL.md lines 45-78, trim to < 2000 tokens
── 2. [CC-2026-W17-hooks-audit-log] hooks audit log (priority=76) ── ...
πŸ’‘ Deep-dive suggestion (you choose to run):
  WebFetch https://anthropic.com/... "verify still applies after 2026-04"
Confirm? [A] All [1,2] Select [N] Skip [D 1] Full diff

* Example only β€” not real project data

What LoreJump does behind the scenes

Your skill just calls lorejump-mcp β€” the scanning, source selection, and entry quality control are continuously maintained by LoreJump.

Information sources

Anthropic official changelog / engineering blog, MCP spec, curated community harnesses (awesome-claude-code, claudekit, agnix, etc.). The full list is maintained internally and iterates over time.

Quality control

Every practice carries applicable dimensions, effort / impact tags, verification date, and a pool of further-reading URLs. Outdated / superseded practices are auto-downgraded.

Beta notes

Run a diagnosis first

Establish your harness baseline score
Install lorejump-mcp

Required prerequisite for /lorejump-harness