From transcend-harness
Critique an existing .claude/ harness — drift from a transcend manifest, missing pillars, over/under-enforcement, stale handoffs, oversized CLAUDE.md, broken imports, and unwired specialized workflows — propose diff-style improvements, and optionally apply the safe ones. Use to review or improve a project's Claude Code harness (whether transcend generated it or not).
How this skill is triggered — by the user, by Claude, or both
Slash command
/transcend-harness:transcend-auditThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
Inspect the target project's `.claude/` harness and produce a critique with
Inspect the target project's .claude/ harness and produce a critique with
concrete, diff-style suggestions. Read-only by default — write only via the
safe-apply flow in Step 3 after explicit confirmation, and never overwrite a
hand-edited or untracked file at all (propose the change instead).
TRANSCEND_ROOT="${CLAUDE_PLUGIN_ROOT:-$(cd "$(dirname "$CLAUDE_SKILL_DIR")/.." && pwd)}"
PROJECT_DIR="${CLAUDE_PROJECT_DIR:-$(pwd)}"
C="$PROJECT_DIR/.claude"
echo "TRANSCEND_ROOT=$TRANSCEND_ROOT"
echo "PROJECT_DIR=$PROJECT_DIR"
echo "--- tree ---"; ( cd "$PROJECT_DIR" && find .claude -maxdepth 3 -type f 2>/dev/null | sort )
echo "--- drift report ---"; sh "$TRANSCEND_ROOT/core/audit/verify-manifest.sh" "$PROJECT_DIR"
echo "--- manifest ---"; cat "$C/.transcend/manifest.json" 2>/dev/null || echo "NO_MANIFEST (harness not transcend-generated, or hand-built)"
echo "--- CLAUDE.md line count ---"; wc -l "$C/CLAUDE.md" 2>/dev/null || echo "NO_CLAUDE_MD"
echo "--- handoff status ---"; sed -n '1,6p' "$C/handoffs/current.md" 2>/dev/null || echo "NO_HANDOFF"
echo "--- settings hooks present? ---"; grep -l '"hooks"' "$C/settings.json" 2>/dev/null || echo "no hooks block"
You may delegate the deeper read-only inspection to the transcend-auditor agent
(context: fork) to keep the main context clean — pass it TRANSCEND_ROOT and
PROJECT_DIR; it returns findings JSON with optional apply plans.
For each, produce findings with a severity (info / warn / error):
modified =
hand-edited: flag as info, mark "preserve — suggest only". missing =
recorded file deleted (warn — offer to regenerate). untracked = present
but unrecorded (never touch; mention only if it conflicts with the harness).
handoffs/current.md reporting modified is the handoff loop working — do
NOT flag it.rules/*.md or a
CLAUDE.md section each). Flag gaps; missing context/handoff is warn ("you're
losing session state").appetite vs actual
hooks. Strict appetite but no blocking hooks = under (warn). A blocking hook
that could fire on routine edits (broad matcher, no if:) = over (warn).
Hooks referencing scripts that don't exist = error.current.md with status: done or an
old updated while work is clearly ongoing; a SessionStart load hook present
but current.md missing; archive bloat.warn); @imports deeper
than 4; dangling @paths; rule paths: globs that match no files.${CLAUDE_PLUGIN_ROOT}/${CLAUDE_PROJECT_DIR}.core/catalog/catalog.yaml) that aren't wired.Group findings by pillar with severities. For each actionable finding, show a diff-style suggestion (the exact rule/settings edit). Partition into:
apply plan whose target is, per
the drift report, a new path (create) or pristine (ok → regenerate /
append / settings-merge).modified (hand-edited) or untracked
file, plus judgment calls (tier changes, restructuring). These are shown as
diffs for the developer to apply by hand.Do not write anything yet.
If there are no safe-apply findings, end after the report. Otherwise ask (AskUserQuestion): [Just report / Apply N safe fixes / Show full diffs first] — after showing diffs, ask again whether to apply.
On "apply", delegate to the transcend-generator agent in merge mode: pass
TRANSCEND_ROOT, PROJECT_DIR, the drift report, and the list of apply plans.
Merge-mode rules (the generator re-verifies each one before writing):
create — write a file that exists neither on disk nor in the manifest.regenerate / append — only if the target's CURRENT sha256 still matches
the manifest (re-hash at write time; the audit may be minutes old).settings-merge — additive JSON merge into a pristine settings.json: append
hook entries / permission lines, never remove or reorder existing ones. Copy
any referenced scripts into .claude/scripts/transcend/ (with lib/common.sh)
and chmod +x.apply plan
slipped through — the generator downgrades it to a suggestion..transcend/manifest.json: refresh/add
files[] hashes for every written file and stamp top-level last_merge
(ISO-8601 UTC).Print the generator's summary — written / skipped (with reasons) / remaining
suggestions — and recommend reviewing with git diff .claude before committing.
npx claudepluginhub ldippo/transcend-harness --plugin transcend-harnessGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.