Task Loop

Task Loop

Scaffold a self-contained loop directory from an existing backlog. The directory contains everything a fresh agent session needs to pick up a task, execute it, and hand off to the next iteration - with no session memory between runs.

When to use

Invoke when the user has a backlog (from /task-plan, hand-written, or any structured task list) and wants to run it through an automated agent loop. The output is a directory with PROMPT.md, run-log.md, and a shared .gitignore.

Process

1. Locate the backlog

Accept a backlog file path. If not provided, look for:

• A file the user recently created or discussed
• Files matching TASKS/*/backlog.md or *TASKS*.md
• Any markdown file with checkbox tasks

Ask if unclear.

2. Read and understand the backlog

Read the backlog to understand:

• Task shape (what fields each task has)
• Verification patterns (global rules, per-task criteria, referenced docs)
• Dependencies between tasks
• Any reference material mentioned in the backlog header

This understanding drives how PROMPT.md is adapted.

3. Ask about execution context

Ask the user:

• "Are there docs or skills that will help execute these tasks?" (e.g. Figma workflow, testing guide, API docs)
• "Any project-specific verification rules beyond what's in the backlog?"
• "What timeout per iteration makes sense?" (helps with the suggested run command)

4. Generate the loop directory

If the backlog is already in a TASKS/<name>/ directory, generate files alongside it. If not, ask where to create the directory or suggest TASKS/<name>/ based on the backlog content.

Generate these files:

PROMPT.md

The loop contract. Self-contained - the agent reads this one file and has the complete protocol. Built from the core protocol template (read references/loop-protocol.md) plus task-specific adaptations:

• File paths - point to this directory's loop-state.md, run-log.md, and backlog.md
• Loop completion token - emit __PROMISE_RL_DONE__ as a standalone final line when no unchecked tasks remain so the default rl promise-token handling can stop cleanly. This is the Ralph-loop "completion promise" expressed as a plain token.
• Verification rules - extracted from the backlog's global and per-task verification patterns. Reference external docs if the backlog mentions them
• Dependency handling - if the backlog has a dependency graph, add instructions to respect it when picking the next task
• Reference docs - if the user mentioned helpful docs or skills, add them to the prompt so the agent knows where to look

The prompt must start with a level-1 heading (markdown linter requirement).

run-log.md

Empty file with # Run Log header. Entries are appended by the agent during execution - one entry per completed or blocked task.

Entry format:

## <ISO-timestamp> | <task-id> | <done|blocked>
- **Commit:** <sha>
- **Verification:** <what was run>
- **Surprises:** <anything unexpected, or "none">

TASKS/.gitignore

If a TASKS/.gitignore doesn't already exist in the parent directory, create one containing loop-state.md. This keeps ephemeral state out of version control while allowing clean deletion of the entire task directory.

5. Present the output

Show the user:

• The generated directory structure
• The PROMPT.md content (or a summary if long)
• How to run it, e.g.: rl <n> -t 30m -- cys "Read and follow TASKS/<name>/PROMPT.md"

Use cys (claude yolo stream: skip-permissions + -p baked in), not bare claude -p - interactive-permission claude stalls the loop the first time it hits a permission prompt with nobody watching.

PROMPT.md anatomy

Every generated PROMPT.md has these sections:

1. Heading - # Prompt
2. Preamble - "You are running in an automated loop. Read these files before doing anything else:" followed by the three file paths
3. Protocol - the loop state machine; the canonical version lives in references/loop-protocol.md
4. Verification - rules adapted from the backlog, referencing external docs where appropriate
5. Completion rule - the four conditions that must all be true before marking a task done
6. Surprises - instructions to capture unexpected findings

Read references/loop-protocol.md for the core template. Adapt it - don't copy it verbatim. Each backlog has different verification needs and reference material.

loop-state.md

Created at runtime by the agent, not pre-created by this skill. The agent creates it when claiming a task. Format:

---
current_task: <task-id>
status: in_progress | verifying | done | blocked
last_commit: ""
next_task: <task-id>
blockers: ""
---

## Checklist

- [ ] Code changes made
- [ ] Verification passed
- [ ] Backlog checkbox updated
- [ ] Commit created

## Surprises

- (none yet)

Gitignored via TASKS/.gitignore - changes every iteration, noisy in version control. Clean deletion of the task directory removes it too.