zrosenfield/ralph-loop

RALPH Loop — Relentless Agentic Loop for Pursued Hardening

Overview

RALPH is an iterative execution pattern where Claude pursues a goal through structured cycles, self-evaluating after each iteration and only stopping when it is genuinely confident the goal has been fully achieved — or when it has exhausted its configured iteration budget.

The loop is designed to prevent the common failure mode of "close enough" — where an agent produces a reasonable first attempt and stops, leaving gaps, inconsistencies, or missed requirements that a more careful pass would catch.

The RALPH Cycle

Each iteration follows five phases:

┌─────────────────────────────────────────────────┐
│  R  →  A  →  L  →  P  →  H                     │
│  │     │     │     │     │                      │
│  │     │     │     │     └─ HARDEN              │
│  │     │     │     │        Decide: DONE or     │
│  │     │     │     │        loop back to R      │
│  │     │     │     │                            │
│  │     │     │     └─── PROBE                   │
│  │     │     │          Find specific gaps,     │
│  │     │     │          weaknesses, omissions   │
│  │     │     │                                  │
│  │     │     └───────── LOOK                    │
│  │     │                Examine output against  │
│  │     │                success criteria        │
│  │     │                                        │
│  │     └─────────────── ACT                     │
│  │                      Execute the plan,       │
│  │                      produce/refine output   │
│  │                                              │
│  └───────────────────── REASON                  │
│                         Decompose goal,         │
│                         define success criteria, │
│                         plan approach            │
└─────────────────────────────────────────────────┘

Phase Details

R — REASON (Iteration 1: Goal Decomposition | Iteration 2+: Gap-Targeted Replanning)

First iteration:

• Parse the user's stated goal into a clear, unambiguous objective
• Decompose the goal into success criteria — concrete, evaluable conditions that must ALL be true for the goal to be considered achieved
• Each criterion should be specific enough to score on a 1–10 scale
• Plan the execution approach (tools, sequence, dependencies)
• Identify what "done" looks like — the acceptance threshold

Subsequent iterations:

• Review the gaps identified in the previous PROBE phase
• Replan specifically to address those gaps
• Do NOT restart from scratch — build on what's already been produced
• Prioritize the lowest-scoring criteria from the previous evaluation

Output a structured plan:

ITERATION [N] — REASON PHASE
=============================
Goal: [restated goal]
Success Criteria:
  1. [criterion] — [what "10/10" looks like]
  2. [criterion] — [what "10/10" looks like]
  ...
Plan: [what this iteration will do]
Focus Areas: [gaps from previous iteration, or "initial execution" if iteration 1]

A — ACT (Execute the Plan)

• Execute the plan from the REASON phase
• Use whatever tools are needed (file creation, web search, computation, etc.)
• Produce tangible output (files, content, data, analysis)
• On iteration 2+, modify/enhance existing output rather than regenerating from scratch (unless the PROBE phase determined a fundamental approach change is needed)

L — LOOK (Evaluate Against Criteria)

• Examine the output produced in the ACT phase
• Score EACH success criterion on a 1–10 scale
• Be brutally honest — this is where most agents fail by being too generous with themselves
• A score of 10 means "I would bet my reputation that this criterion is perfectly met"
• A score of 7-8 means "solid but there's room for improvement"
• A score below 7 means "this needs more work"

Output a structured evaluation:

ITERATION [N] — LOOK PHASE
============================
Criterion Scores:
  1. [criterion]: [score]/10 — [brief justification]
  2. [criterion]: [score]/10 — [brief justification]
  ...
Overall Confidence: [weighted average]/10

Scoring calibration guide:

• 10/10: Flawless. No possible improvement. Would pass expert review.
• 9/10: Excellent. Minor stylistic preferences might differ but substance is complete.
• 8/10: Very good. One small gap or imperfection that most people wouldn't notice.
• 7/10: Good. Meets the requirement but has a noticeable weakness.
• 6/10: Acceptable. Gets the job done but clearly could be better.
• 5/10 or below: Needs significant work. Do not consider stopping.

P — PROBE (Find the Gaps)

• For any criterion scoring below the completion threshold, identify EXACTLY what's missing or weak
• Be specific — "the introduction could be better" is useless; "the introduction doesn't establish the business context for the recommendation" is actionable
• Consider edge cases, unstated requirements, and quality dimensions the criteria might not fully capture
• Ask: "If I handed this to the user right now, what would they push back on?"

Output a structured gap analysis:

ITERATION [N] — PROBE PHASE
=============================
Gaps Found:
  - [Criterion N]: [specific, actionable description of what's missing]
  - [Criterion M]: [specific, actionable description of what's weak]
  ...
Unstated Issues: [anything not captured by criteria but still problematic]
Verdict: [CONTINUE | COMPLETE]

H — HARDEN (Decide: Done or Iterate)

Apply the completion rules (see below) to determine whether to:

1. COMPLETE — Declare the goal achieved, present final output to the user
2. ITERATE — Feed the gaps from PROBE back into the next REASON phase

If completing: Present the output with a brief summary of the iterations taken and final confidence scores. If iterating: Proceed immediately to the next REASON phase with the gap analysis as input.

---

Completion Rules

The loop terminates when ANY of these conditions are met:

1. Confidence Threshold Reached

• ALL criteria score at or above the completion threshold
• Default threshold: 9/10
• Configurable per goal (see Configuration section)

2. Max Iterations Reached

• The loop has executed the maximum allowed iterations
• Default: 5 iterations
• Configurable per goal
• When hitting the cap: Present the best output achieved with honest assessment of remaining gaps

3. Diminishing Returns Detected

• The overall confidence score improved by less than 0.5 points between iterations
• AND the loop has run at least 2 iterations
• This prevents infinite loops on goals where perfection isn't achievable
• When triggered: Explain to the user what's plateauing and why

4. Fundamental Blocker Identified

• The PROBE phase identifies an issue that cannot be resolved within the current context (e.g., missing data, tool limitations, ambiguous requirements needing user input)
• When triggered: Present current best output and clearly state what's blocking further progress

---

Configuration

RALPH can be configured per goal by specifying parameters at the start:

| Parameter | Default | Description | |-----------|---------|-------------| | max_iterations | 5 | Hard cap on iteration count | | completion_threshold | 9 | Minimum score (1-10) ALL criteria must reach | | strictness | high | low (7+), medium (8+), high (9+), perfectionist (10) | | verbose | true | Whether to output phase logs (set false for cleaner output) |

To configure, the user can say things like:

• "Use RALPH with max 3 iterations" → max_iterations=3
• "RALPH this but keep it quick" → max_iterations=3, strictness=medium
• "RALPH this to perfection" → strictness=perfectionist, max_iterations=8
• "RALPH this" → all defaults

---

Usage Pattern

When the user invokes RALPH (or when the skill determines RALPH is appropriate), follow this sequence:

Step 1: Acknowledge and Configure

Starting RALPH loop.
Goal: [user's goal]
Configuration: [max_iterations] iterations | threshold: [completion_threshold]/10

Step 2: Execute the Loop

Run RALPH cycles. On each iteration, output the phase logs (if verbose=true) so the user can see the agent's reasoning.

Step 3: Present Results

When complete, present:

• The final output/deliverable
• A completion summary:

RALPH COMPLETE
==============
Iterations: [N] of [max]
Final Confidence: [score]/10
Criteria Met: [N]/[total] at threshold
Termination Reason: [threshold reached | max iterations | diminishing returns | blocker]

---

Anti-Patterns to Avoid

1. Generous self-scoring: The most common failure. If in doubt, score LOWER. A 9 should mean you'd genuinely be surprised if the user found a problem with that criterion.

2. Cosmetic iterations: Don't waste iterations on formatting tweaks when substance is lacking. Prioritize the lowest-scoring criteria.

3. Scope creep: Each iteration should address identified gaps, not invent new requirements. Stick to the success criteria defined in iteration 1 (though PROBE can surface unstated issues that are genuinely important).

4. Restarting from scratch: Unless PROBE identifies a fundamental approach problem, iterate on existing output rather than regenerating. Build, don't rebuild.

5. Ignoring diminishing returns: If the score barely moved, don't keep grinding. Present what you have and explain the plateau.

---

Examples of RALPH in Action

Example 1: Document Generation

User: "RALPH: Create a project status report for the SharePoint AI initiative"

RALPH would:
- R: Define criteria (executive summary clarity, data accuracy, completeness of milestones, actionable next steps, appropriate tone)
- A: Generate the report
- L: Score each criterion honestly
- P: "Executive summary buries the lead — revenue impact should be sentence 1, not paragraph 3"
- H: Continue → feed gap back
- R2: Replan to restructure executive summary
- A2: Revise the report
- L2: Re-score — all criteria now 9+
- H2: COMPLETE

Example 2: Data Extraction

User: "RALPH: Extract all recipes from these uploaded documents and validate completeness"

RALPH would:
- R: Define criteria (all recipes found, ingredients complete, instructions parseable, measurements standardized, edge cases handled)
- A: Extract recipes
- L: Score — finds 3 recipes missed from a multi-column page
- P: "Multi-column layout caused parser to skip items. Also, 2 recipes have vague measurements like 'a pinch'"
- H: Continue
- ... iterate until extraction is validated

Example 3: Analysis Task

User: "RALPH with max 3 iterations: Analyze our competitive landscape in the AI agent space"

RALPH would:
- R: Define criteria (market coverage, competitor accuracy, strategic insight depth, actionable recommendations)
- A: Research and produce analysis
- L: Score honestly
- P: Find gaps in coverage or depth
- H: Iterate up to 3 times, then present best result with honest gap assessment

---

Integration Notes

• RALPH works with any other skill — it's a meta-pattern that wraps around execution
• The ACT phase can invoke any tools or skills needed
• RALPH logs can be saved to files for audit trails if needed
• The pattern works for both simple tasks (2 iterations) and complex tasks (5+ iterations)
• When used inside SharePoint workflows, RALPH can leverage SPO tools during the ACT phase and validate against SharePoint-specific quality criteria