paulrberg/loop-skill

Loop Skill

Run another agent skill multiple times in sequence, reporting only the final net result.

Arguments

• Target skill invocation: required. Accept $skill-name, skill-name, a path to SKILL.md, or a natural-language instruction that names the skill and its arguments.
• --runs <n> or -n <n>: optional positive integer iteration count. Default: 3.
• Additional target arguments: forward unchanged to the target skill on every iteration.

Preconditions

• Do not run in Plan mode. If the current mode, tools, or user instruction indicate planning-only work, stop and say: $loop-skill is execution-only and cannot run in Plan mode. Switch to execution/default mode and invoke it again.
• Require a target skill. If the request does not identify one, stop and ask for the skill invocation to repeat.
• Resolve the target skill before starting:
  • For $skill-name or skill-name, read ../skill-name/SKILL.md.
  • For a path, read that SKILL.md.
  • For an already-loaded skill block in the conversation, use that block only when it clearly matches the requested target.
  • If the target cannot be resolved, stop and report the missing skill.
• Refuse to target loop-skill directly.
• Refuse targets whose required workflow commits, tags, stashes, force-resets, or otherwise depends on intermediate git history. This loop makes no commits.

Workflow

1. Initialize

1. Parse the run count. Use 3 unless the user explicitly asks for a different positive integer.
2. Read the target skill instructions once.
3. Verify repository context with git rev-parse --git-dir when the target may inspect diffs. If it fails, continue only if the target skill is explicitly non-repo work.
4. Snapshot the initial state for the final net report:
   • git status --short
   • git diff --stat
   • git diff --name-only --diff-filter=ACMR
   • git ls-files --others --exclude-standard

2. Iterate

For each iteration from 1 through the requested run count:

1. Run the target skill inline as if invoked with the user's target arguments.
2. Override the target skill with these loop constraints:
   • Do not create commits, tags, branches, stashes, or other persistent git history.
   • Do not produce a full user-facing report unless the target must stop or ask for direction.
   • Keep scope stable unless the target's own instructions require a narrower discovered scope.
3. If the target skill hits a stop condition, stop the loop and produce the net report for completed work plus the blocking condition.

3. Verify

After the final completed iteration, verify the net final state with the narrowest command that proves the touched behavior.

• Prefer project recipes such as just test, just lint, just mdformat-check, or a narrower recipe when present.
• Name skipped checks and why.

4. Report

Produce one final report only. Base it on the net diff from the initial snapshot to the final state, not on intermediate fixes that were later changed again.

Use these headings in order, omitting sections that do not apply:

Loop

Target skill, target arguments, requested iterations, completed iterations, and whether the loop stopped early.

Net Changes

Final files and behaviors changed. Mention intermediate work only when it explains a final decision or residual risk.

Verification

Commands run and outcomes, including skipped checks.

Residual Risks

One line per risk: Assumed <assumption>; if wrong, <what breaks>; check via <command or inspection>. Write None. when there are none.