dbmrq/agent-skills/ralph-loop

Ralph Loop

When to use

• User wants to run a Ralph loop against an existing plan
• User wants to write or improve a ralph-loop.sh (or equivalent) script
• User is inside a loop run (prompt mentions automated Ralph loop)
• User asks how Ralph loops work

If there is no plan yet, create one first with the ralph-loop-plan skill, then return here for the script and orchestration.

What a Ralph loop is

A Ralph loop runs a coding agent repeatedly with a fresh context each time, not one long chat.

Core insight: the agent does not carry the whole project in memory. The filesystem and git do. Each run reads current state, does one checklist item, verifies, updates the plan, exits. The script starts the next run.

Named after Geoffrey Huntley's formulation: "Ralph is a Bash loop" — while + agent CLI + prompt file.

End-to-end workflow

When the user says "use a Ralph loop to build/create X" (any stack — app, library, CLI, service):

1. Plan — Use ralph-loop-plan to write docs/<feature>-plan.md (or user-chosen path).
2. Script — Generate scripts/ralph-loop.sh from script-template.sh; point --plan-file at the plan.
3. Dry run — ./scripts/ralph-loop.sh --dry-run to inspect the first prompt.
4. Run — ./scripts/ralph-loop.sh (optionally tune --max-runs, --max-no-progress-runs).
5. Human gate — After all checkboxes are done, user runs the plan's non-checkbox manual validation section.

Per-run agent prompt (canonical)

Use this shape. Adapt paths and repo name; keep the rules intact.

You are running inside an automated Ralph loop.
 
Goal:
Work on the next unchecked item in `<PLAN_FILE>`, verify it, then mark it done only if verification actually passed.
 
Operating rules:
-Read `<PLAN_FILE>` first. It is the source of truth for scope and ordering.
-Work on **only** the next unchecked checklist item. Do not start later items.
-Keep changes scoped to that item; follow existing project patterns.
-Add or update focused tests when appropriate for this item.
-Run the fastest relevant verification (unit tests, lint, build) and record the exact command in your summary.
-When complete, change that item's checkbox from `[ ]` to `[x]`.
-If you discover constraints, commands, file paths, or dependencies that future agents would struggle to rediscover, add a **Notes for later items:** block under the completed item or on relevant future unchecked items. Do not rewrite unrelated plan sections.
-If blocked or unsafe to complete, document the blocker under that item, leave it unchecked, and stop.
-Do not create git commits unless the loop script explicitly tells you to; many loops commit after each run automatically.
-Do not expand scope, refactor unrelated code, or reorder checklist items.
 
Before finishing, report: what changed, what verification ran (with result), whether the item was marked complete, and any plan notes you added.

See prompt-template.md for a fill-in version with run metadata.

Guardrails (always enforce)

These belong in both the script and the per-run prompt.

Hard stops (script)

Soft rules (prompt)

Optional extras

• `--dry-run` — Print first prompt; no agent call.
• Completion promise — Agent emits <promise>COMPLETE</promise> when plan is empty; script validates. Useful for hook-based loops; checklist counting is often enough for shell loops.
• Baseline test run before loop — Ensures agents don't inherit an already-red tree (document in plan item 1).

Improving the plan during loops

Agents should enrich the plan; they must not destabilize it.

Allowed (encourage):

• **Notes for later items:** under a completed item — implementation discoveries, test commands, gotchas, file paths
• Short additions on future unchecked items — verification command, dependency ("requires item 7 first"), known pitfall
• Blocker subsection under a stuck item — what failed, what was tried, what human decision is needed

Forbidden:

• Renumbering or reordering items mid-loop
• Deleting unchecked items without user approval
• Marking items done without verification
• Moving human-only work into checkbox items

If an agent finds the plan structure itself is wrong (missing integration phase, item too large), add a Plan amendment note at the bottom and leave a checkbox unchecked with the blocker — do not silently reorganize.

Writing a ralph-loop script

Use script-template.sh as the starting point. Key non-obvious implementation details:

1. Fresh process each iteration

Each run is a new agent process with -p "$prompt". Do not reuse one long session — that defeats the loop.

Detect CLI: agent, cursor-agent, or cursor agent. Honor CURSOR_AGENT_BIN.

2. Progress detection = checklist parsing

Count unchecked items with a stable pattern. Prefer numbered items:

^\s*-\s+\[\s\]\s+\*\*\d+\.

Generic fallback for unnumbered plans:

^\s*-\s+\[\s\]\s+

Expose unchecked_count, next_unchecked_item, and compare before/after each run.

3. Two different "stuck" signals

• No file changes — Agent may have exited without doing work, or work was already done.
• File changes but same unchecked count — Agent worked but did not finish or forgot to check off. Track separately from no-change.

4. Commit strategy

Common pattern: script commits after each run that changed files (ralph-loop: run N). Agent prompt says do not commit to avoid double commits. Pick one owner — script is usually safer.

5. Allowed dirty paths at start

When enforcing clean tree, always allow: the plan file, the loop script, .ralph-loop/, and .gitignore. Pass plan path dynamically — do not hardcode project-specific paths.

6. Pre-flight

git rev-parse --show-toplevel   # must be in repo
test -f "$PLAN_FILE"            # plan exists
test ! -f .git/MERGE_HEAD       # no merge

7. Exit codes

Working inside a loop (agent checklist)

When the prompt says you are in a Ralph loop:

1. Read the full plan (at least through the next unchecked item and any linked spec).
2. Identify the first - [ ] checkbox item.
3. Implement only that item.
4. Run verification listed in the item (or the fastest relevant default for this repo).
5. Mark [x] only if verification passed.
6. Add Notes for later items if you learned something non-obvious.
7. Summarize for the log; exit cleanly.

If blocked: document under the item; leave unchecked; stop.

Script generation checklist

When creating a new loop script for a project:

• [ ] --plan-file defaults to the project's plan path
• [ ] --max-runs, --max-no-change-runs, --max-no-progress-runs with sane defaults (e.g. 24 / 3 / 3)
• [ ] --dry-run, --allow-dirty, --help
• [ ] Agent auto-detection + CURSOR_AGENT_BIN
• [ ] Logs under .ralph-loop/logs/run-NNN.log
• [ ] .ralph-loop/ in .gitignore (logs only; plan stays tracked)
• [ ] Prompt matches canonical shape above
• [ ] chmod +x the script

Related

• Plans and task breakdown → ralph-loop-plan skill
• Detailed theory and alternatives → reference.md
• Prompt copy-paste template → prompt-template.md
• Script starting point → script-template.sh