Ralph Loop Setup

Ralph loop setup

Resolve configuration and seed the loop files. Never start the loop. Setup ends with a summary and an instruction to run /ralph-loop start.

Writing is done by scripts/seed-ralph-loop.sh, not by hand. Your job is to resolve values and call it. Hand-authoring the loop file reintroduces the unsubstituted-placeholder failure the script exists to prevent: a stray {{MAX_ITERATIONS}} in the frontmatter fails the hook's numeric validation and silently deletes the loop.

Interview

Ask only what you cannot resolve yourself. Use structured questions, not prose.

1. Preset. If not given:

• engineering-delivery — drive an epic through implement, review,

validate, and merge request, one task per iteration.

• ad-hoc — repeat a single prompt until it is done.
• custom — define your own steps.

1. Target. The epic slug or ID for engineering delivery; the task prompt

for ad-hoc; the step list for custom.

1. Budgets. Max iterations. Default 50, but for engineering delivery

propose tasks × 6 + 10, since a 12-task epic will not fit in 50.

1. Completion promise. Propose a default and confirm it. For an epic, the

slug upper-snake-cased with _COMPLETE.

1. Environment. Only for presets that need it, per

[references/environment-resolution.md](references/environment-resolution.md).

Workflow

1. Resolve the agent

CLAUDE_PLUGIN_ROOT set means claude; CURSOR_PROJECT_DIR set means cursor. This determines the base directory (.claude/loop or .cursor/loop). There is no pointer file and no --ralph-dir flag.

2. Resolve the preset inputs

engineering-delivery

• Resolve {epic} per

[delivery-conventions.md](../tasks/references/delivery-conventions.md): a slug, an epic ID resolved via the backlog row, or a path.

• Locate tasks.md and design.md. Fail loudly, naming the missing file, if

either is absent.

• Derive a dependency-safe task order: topological by declared dependencies,

stable by document order on ties. Render as N. {TASK_ID} — (depends on: ).

• Resolve the branch. Report the expected branch; never create or switch one.
• Resolve validation commands, tracker actions, and UI signals per

[references/environment-resolution.md](references/environment-resolution.md).

ad-hoc

Write the task prompt to a file and pass --prompt-file. Apply [../ralph-loop/references/prompt-authoring.md](../ralph-loop/references/prompt-authoring.md): explicit completion criteria, a verification step each iteration, and an escape hatch for being stuck.

custom

Write the step definitions to a file and pass --steps-file. Each step needs a name, what to do, and which step comes next. See [../ralph-loop/references/preset-authoring.md](../ralph-loop/references/preset-authoring.md).

3. Seed

Call the script. Every template value goes through --set:

scripts/seed-ralph-loop.sh \
  --agent claude \
  --preset engineering-delivery \
  --run-id "{epic}-$(date -u +%Y%m%d-%H%M%S)" \
  --max-iterations 70 \
  --completion-promise CHECKOUT_FOUNDATION_COMPLETE \
  --session-id "$SESSION_ID" \
  --set EPIC=checkout-foundation \
  --set BRANCH=feat/checkout-foundation \
  --set TASKS_PATH=docs/work/checkout-foundation/tasks.md \
  --set DESIGN_PATH=docs/work/checkout-foundation/design.md \
  --set FIRST_ITEM=CHK01-01 \
  --set "WORK_SEQUENCE=$(cat sequence.txt)" \
  --set "GOAL=..." --set "DONE_CRITERIA=..." --set "PRESET_CONTEXT=..."

Run with --dry-run first when anything is uncertain. The script refuses to overwrite a loop past iteration 1 without --force, and exits non-zero on any unresolved placeholder.

4. Report

Files written, resolved configuration, expected branch, and "Run /ralph-loop start to begin." Nothing else.

Policies

• MUST NOT execute loop steps or launch sub-agents. Setup only writes files.
• MUST NOT create or switch git branches.
• MUST NOT hand-author active.md, loop-state.md, or context.md.
• MUST fail loudly, naming the file, when a required source is missing.
• MUST NOT invent a task order that ignores declared dependencies.

Anti-patterns

• Writing the loop file directly instead of calling the seed script.
• Starting the loop after seeding it.
• Guessing at validation commands rather than resolving them from the repo.
• Setting a completion promise the loop has no way to verify.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: carinyaparc
• Source: carinyaparc/skills
• License: MIT

Install and usage instructions live in the source repository linked above.