loop-author-a

Action skill — Write session prompts for agent loops following the loop-author guide.

Input

$ARGUMENTS

---

Write SESSION_WORKER.md and SESSION_REVIEWER.md for an agent loop. These prompts define what the loop accomplishes and how the reviewer evaluates completion.

Gather Context

Before writing, read these sources to understand the repo and task:

1. Repo Context

# Project instructions
cat CLAUDE.md 2>/dev/null
cat .claude/CLAUDE.md 2>/dev/null

# Recent work
git log --oneline -20

# Repo structure (top-level)
ls -la

2. Task Context

From $ARGUMENTS, extract:

• What the loop should accomplish
• Where in the codebase the work lives
• Why this work matters (if stated)
• Constraints (language, patterns, dependencies)

3. Loop Author Guide

Follow the session prompt anatomy from agents/prompts/loop-author.md (or ~/claude-tooling/agents-src/prompts/loop-author.md). The four required sections are:

1. Objective — single paragraph, no ambiguity
2. Context — key files, existing patterns, constraints
3. Behavior Targets — observable goals, not implementation steps
4. Acceptance Criteria — verifiable conditions for STOP

Write SESSION_WORKER.md

# Objective

{One paragraph. What is the end goal? What is the output?}

{Read X, Y, Z before starting — point to specific files.}

## Context

{What exists today. Key files and directories. Existing patterns to follow.
What has been tried. Hard constraints.}

## Behavior Targets

When this is done, the following should be true:

1. {Observable behavior — written from a tester's perspective}
2. {Observable behavior}
3. ...

## Acceptance Criteria

- {Verifiable condition — the minimum bar}
- {Verifiable condition}
- ...

## Constraints

- {Hard constraint — language, patterns, dependencies}
- {Hard constraint}

Worker Prompt Principles

• Objective is the north star. If the objective is vague, the loop wanders. Be specific.
• Point to files, not concepts. "Read src/auth/middleware.ts" not "understand the auth system."
• Behavior targets are observable. "Users can log in with SSO" not "Add OAuthProvider class."
• No decomposition. Don't tell the worker how to break down the work — that's what TODO.md is for.
• No code snippets. Session prompts are strategic, not tactical.

Write SESSION_REVIEWER.md

# Review Brief

You are reviewing progress toward: {one-line objective summary}

## What to Verify

- {Verification point — what to check at each iteration}
- {Verification point}

## Quality Bar

- {What good looks like for this specific task}
- {Standards that apply — tests, linting, patterns}

## Watch For

- {Known risk or failure mode}
- {Common mistake for this type of work}

## Done Condition

Write STOP.txt when:
- {Specific, observable condition}
- {Specific, observable condition}

Reviewer Prompt Principles

• Same objective, different perspective. The reviewer sees the same goal but evaluates rather than builds.
• Shorter than the worker prompt. Reviewers need clarity, not volume.
• Concrete done conditions. "All tests pass" not "code is clean."
• Watch-for items are specific. "Over-engineering the error handling" not "bad code."

Output Location

Write the prompts to:

agents/loops/{loop-id}/session/SESSION_WORKER.md
agents/loops/{loop-id}/session/SESSION_REVIEWER.md

Create the directory structure if it doesn't exist:

mkdir -p agents/loops/{loop-id}/session

After Writing

1. Show the user both prompts for review
2. Ask if they want to adjust anything before launch
3. If they're satisfied, suggest launching with loop-launch-a or directly:

   ./agents/run.sh --loop-id {loop-id}
   

Common Mistakes to Avoid

Mistake	Fix
Vague objective	One paragraph, no ambiguity — name the output
Steps instead of goals	Describe what, not how
No acceptance criteria	List verifiable conditions
Reviewer copies worker	Write from reviewer perspective — it's a review brief
Too much context	Only include what's needed to start work
Implementation approach in prompt	Let the worker decide, reviewer course-correct