do-and-judge

Task

Execute a single task by dispatching an implementation sub-agent, verifying with an independent judge, and iterating with feedback until passing or max retries exceeded.

Arguments

Argument	Format	Default	Description
task	Free-form text	Required	Task description to execute
--model	haiku|sonnet|opus	auto-selected	Explicit user override for all sub-agents: implementation, meta-judge, and judge. When omitted, you MUST select the model per the Model Selection Policy — there is no fixed fallback tier. When provided, the user's choice wins over the policy for every sub-agent — see the Escalation Rule for how escalation interacts with an explicit override.
--strict	--strict	false	Disable the Iteration Discretion Rule - the task passes ONLY when score >= 4.0, otherwise retry until max retries is reached.

Example: /do-and-judge Refactor the UserService class to use dependency injection --strict

Context

This command implements a single-task execution pattern with meta-judge → LLM-as-a-judge verification. You (the orchestrator) dispatch a meta-judge (to generate evaluation criteria) and an implementation agent in parallel, then dispatch a judge with the meta-judge's evaluation specification to verify quality. If verification fails, you launch new implementation agent with judge feedback and iterate until passing (score ≥4, or accepted per the Iteration Discretion Rule) or max retries (3) exceeded.

Key benefits:

• Fresh context - Implementation agent works with clean context window
• Structured evaluation - Meta-judge produces tailored rubrics and checklists before judging
• External verification - Judge applies meta-judge specification mechanically — catches blind spots self-critique misses
• Parallel speed - Meta-judge and implementation run simultaneously
• Feedback loop - Retry with specific issues identified by judge
• Quality gate - Work doesn't ship until it meets threshold

CRITICAL: You are the orchestrator only - you MUST NOT perform the task yourself. IF you read, write or run bash tools you failed task imidiatly. It is single most critical criteria for you. If you used anyting except sub-agents you will be killed immediatly!!!! Your role is to:

1. Analyze the task and select the model per the Model Selection Policy — sonnet/haiku by default, opus only when earned
2. Dispatch meta-judge AND implementation agent in parallel as foreground agents (meta-judge first in dispatch order)
3. Dispatch judge agent with meta-judge's evaluation specification
4. Parse verdict and iterate if needed (max 3 retries)
5. Report final results or escalate

RED FLAGS - Never Do These

NEVER:

• Read implementation files to understand code details (let sub-agents do this)
• Write code or make changes to source files directly
• Skip judge verification to "save time"
• Read judge reports in full (only parse structured headers)
• Proceed after max retries without user decision

ALWAYS:

• Use Task tool to dispatch sub-agents for ALL implementation work
• Dispatch meta-judge and implementation agent in parallel (meta-judge FIRST in dispatch order)
• Wait for BOTH meta-judge and implementation to complete before dispatching judge
• Pass meta-judge evaluation specification to the judge agent
• Include CLAUDE_PLUGIN_ROOT=${CLAUDE_PLUGIN_ROOT}`` in prompts to meta-judge and judge agents
• Parse only VERDICT/SCORE/ISSUES from judge output
• Iterate with feedback if verification fails

Model Selection Policy

Picking the model is the single highest-leverage decision you make — more than any prompt wording, it decides whether the task comes back correct and how long it takes. You MUST NOT treat it as a formality: name the tier and give a one-line justification before dispatching. Reaching for the strongest model because you did not want to think is a failure, not caution.

Tier default: sonnet and haiku are the default. opus is reserved and opt-in — it MUST be earned by a trigger in the table below, never picked because you are unsure.

Selection Rules

Task shape	Tier	Examples
Single documentation/text file correction — no code, no cross-file reasoning	haiku	Fix a typo, update a link, correct a stale command in a README
Small, few-line (~10 lines or fewer), mechanical code change confined to one file	haiku	Bump a constant, add a guard clause, rename a local, edit a config value
Code writing — new functions, components or tests, single-module changes, established patterns	sonnet	Add an endpoint, write a service method plus tests, refactor one module
Multi-file refactoring (~3+ files, or any file count when a shared contract changes) OR critical (auth, payments/billing, data integrity, irreversible migration, public API break) OR complex logic (concurrency, non-trivial algorithms, architectural decisions)	opus	Cross-cutting refactor, auth or payment logic, schema migration, novel algorithm design

Precedence (MANDATORY): evaluate EVERY row, not just the first that matches. When more than one row matches, the HIGHEST matching tier wins — criticality and complexity always override size. A four-line null check inside a security-critical auth handler matches both the haiku row and the opus row, and is therefore opus. The critical list is exhaustive, not illustrative: shipping to production, touching real users, or adding to a public API are NOT triggers, so a new endpoint with validation in one service file stays sonnet. Mechanical-breadth carve-out: breadth alone is not complexity — a purely mechanical change (e.g., renaming a symbol across many files, with no logic or contract change) stays at the tier its content earns no matter how many files it touches, so mechanically renaming a symbol across 40 files is haiku or sonnet work, not opus; this carve-out does NOT cover a shared-contract change (already an opus trigger above), so extracting a shared interface across files remains opus.

Tie-breaker: ONLY when no row matches cleanly — the task sits genuinely between two tiers — pick the cheaper tier. You MUST NOT bias up to opus to hedge; the Escalation Rule makes a cheap first guess recoverable, and one recovered run costs far less than over-provisioning every run.

Role Pairing

Any model-assigned pipeline has up to three roles — producer (does the work), criteria-setter (defines what "correct" means), evaluator (checks the work against those criteria); in this skill they instantiate as implementation / meta-judge / judge. Default: the SAME tier for all roles — and where a pipeline has no separate criteria-setter (e.g. a plan step or stage simply assigned a model), this default is the whole rule.

Only for a non-obvious task you MAY raise the criteria-setter alone by one tier, so the criteria are sharper than the work being evaluated. Non-obvious is testable: the tier was decided by the Tie-breaker (no Selection Rules row matched cleanly), OR the task states no checkable acceptance condition.

Pattern	Criteria-setter (meta-judge)	Producer + evaluator (implementation + judge)	Use when
Sharpened-haiku	sonnet	haiku	The work is trivial, but what counts as "correct" is not obvious
Sharpened-sonnet	opus	sonnet	Code work with ambiguous or high-consequence acceptance criteria that does not itself hit an opus trigger

Producer and evaluator MAY be a differnt tier. You MAY decide to raise the evaluator alone if criteria list produced by criteria-setter looks too complex, but you MUST NOT set the criteria-setter below the producer tier.

Escalation Rule

Bump BOTH producer and evaluator (implementation and judge) one tier for the next iteration when either trigger fires:

1. Low first-iteration quality — a low score, or issues showing the model misunderstood the task rather than merely missing details.
2. The user complains that quality is too low or the results are wrong — at any point, including after a reported PASS.

Ladder: haiku → sonnet → opus. opus is the ceiling — there is no further tier. If opus-tier work still fails, escalate to the user, never loop.

• Explicit --model carve-out (the ONLY statement of this rule): an explicit --model is a user override, so trigger (1) MUST NOT silently overrule it — continue iterate with override model till you reach max retry limit. If target still not meet at the end, highlight the found issues and propose to the bump to user. Trigger (2) IS that approval, so it bumps immediately.
• Escalation moves producer and evaluator only. A criteria-setter that already produced the evaluation specification is NOT re-run and NOT re-tiered — changing the criteria mid-task invalidates the comparison across attempts.
• Escalation is a complement to, never a substitute for, a genuine root-cause fix. You MUST still pass the judge's specific feedback into the retry; re-dispatching the same prompt at a higher tier and hoping is prohibited.
• Escalation is orthogonal to the score thresholds and the Iteration Discretion Rule — it changes which model runs the next iteration, never whether an iteration is warranted.

Cross-Provider Equivalence

When this skill runs outside the Anthropic model context, map the tier to the nearest model of the same class:

Tier	Role	Comparable models from other providers
haiku	Fast and cheap; mechanical work	gemini-flash-lite, gemma class, gpt-oss class, small open-weight models
sonnet	Balanced workhorse; most code writing	gemini-pro class and full gemini-flash (not the -lite variant, which is haiku-tier), GPT-5-mini class, large Qwen / DeepSeek class
opus	Frontier reasoning; critical or complex work	whatever the provider sells as its extended / deliberate-reasoning tier — currently GPT-5.5, deep-think modes, Kimi K3 class, any model whose advantage is longer deliberation rather than throughput

The mapping is by capability tier, not by name — exact names drift as vendors ship new models. Every rule above is expressed in tiers, so on another provider: map tier → your model of that class, then apply the selection, pairing and escalation rules unchanged.

Process

Phase 1: Task Analysis and Model Selection

Resolve configuration first: STRICT_MODE = --strict present || false. Strip all flags from the task text — never pass them into sub-agent prompts.

Unless the user passed --model, assess the task on three axes, then read the tier straight off the Selection Rules table:

• Scope — one file, one module, or multiple files?
• Complexity — mechanical edit, established pattern, or novel/intricate logic?
• Risk — isolated and reversible, internal, or critical per the exhaustive list in the Selection Rules opus row?

State the three findings, the chosen tier, and a one-line justification before dispatching. Then apply Role Pairing to decide the meta-judge tier — same tier as implementation unless the task is genuinely non-obvious. If the user passed --model, neither step runs: that one tier is used for implementation, meta-judge and judge alike, and Role Pairing MUST NOT raise the meta-judge above it.

Specialized Agents: Common agents from the sdd plugin include: sdd:developer, sdd:researcher, sdd:software-architect, sdd:tech-lead, sdd:business-analyst. If the appropriate specialized agent is not available, fallback to a general agent without specialization. You MUST use general-purpose every time, when there no direct coralation between task and specialized agent, or agent is not available!

Phase 2: Dispatch Meta-Judge and Implementation Agent (IN PARALLEL)

CRITICAL: Launch BOTH agents in a single message using two Task tool calls. The meta-judge MUST be the first tool call in the message so it can observe artifacts before the implementation agent modifies them.

Both agents run as foreground agents. Wait for both to complete before proceeding to Phase 3.

2.1 Meta-Judge Prompt

The meta-judge generates an evaluation specification (rubrics, checklist, scoring criteria) tailored to this specific task. It will return to you the evaluation specification YAML.

## Task

Generate an evaluation specification yaml for the following task. You will produce rubrics, checklists, and scoring criteria that a judge agent will use to evaluate the implementation artifact.

CLAUDE_PLUGIN_ROOT=`${CLAUDE_PLUGIN_ROOT}`

## User Prompt
{Original task description from user}

## Context
{Any relevant codebase context, file paths, constraints}

## Artifact Type
{code | documentation | configuration | etc.}

## Instructions
Return only the final evaluation specification YAML in your response.

Use Task tool:
  - description: "Meta-judge: {brief task summary}"
  - prompt: {meta-judge prompt}
  - model: {meta-judge model — the user's `--model` if one was passed; otherwise same as implementation, or one tier up per Role Pairing}
  - subagent_type: "sadd:meta-judge"

2.2 Implementation Agent Prompt

Construct the implementation prompt with these mandatory components:

Zero-shot Chain-of-Thought Prefix (REQUIRED - MUST BE FIRST)

## Reasoning Approach

Before taking any action, think through this task systematically.

Let's approach this step by step:

1. "Let me understand what this task requires..."
   - What is the specific objective?
   - What constraints exist?
   - What is the expected outcome?

2. "Let me explore the relevant code..."
   - What files are involved?
   - What patterns exist in the codebase?
   - What dependencies need consideration?

3. "Let me plan my approach..."
   - What specific modifications are needed?
   - What order should I make them?
   - What could go wrong?

4. "Let me verify my approach before implementing..."
   - Does my plan achieve the objective?
   - Am I following existing patterns?
   - Is there a simpler way?

Work through each step explicitly before implementing.

Task Body

## Task
{Task description from user}

## Constraints
- Follow existing code patterns and conventions
- Make minimal changes to achieve the objective
- Do not introduce new dependencies without justification
- Ensure changes are testable
- Critical: you not allowed to use any mutation git commands, including, but not limited: commit, stash, push, checkout, reset, revert, etc. Except cases when task EXPLICITLY allows or requires it. You can use non-mutation git commands, including, but not limited: status, diff, log, branch, etc.

## Output
Provide your implementation along with a "Summary" section containing:
- Files modified (full paths)
- Key changes (3-5 bullet points)
- Any decisions made and rationale
- Potential concerns or follow-up needed

Self-Critique Suffix (REQUIRED - MUST BE LAST)

## Self-Critique Verification (MANDATORY)

Before completing, verify your work. Do not submit unverified changes.

### Verification Questions

| # | Question | Evidence Required |
|---|----------|-------------------|
| 1 | Does my solution address ALL requirements? | [Specific evidence] |
| 2 | Did I follow existing code patterns? | [Pattern examples] |
| 3 | Are there any edge cases I missed? | [Edge case analysis] |
| 4 | Is my solution the simplest approach? | [Alternatives considered] |
| 5 | Would this pass code review? | [Quality check] |

### Answer Each Question with Evidence

Examine your solution and provide specific evidence for each question.

### Revise If Needed

If ANY verification question reveals a gap:
1. **FIX** - Address the specific gap identified
2. **RE-VERIFY** - Confirm the fix resolves the issue
3. **UPDATE** - Update the Summary section

CRITICAL: Do not submit until ALL verification questions have satisfactory answers.

Dispatch

Determine the optimal agent type based on the task and avaiable agents, for exmple: code implementation -> sdd:developer agent. If you not sure, better use general-purpose agent, than dispatch incorrect agent type.

Use Task tool:
  - description: "Implement: {brief task summary}"
  - prompt: {constructed prompt with CoT + task + self-critique}
  - model: {selected implementation model}
  - subagent_type: "{selected agent type}"

2.3 Parallel Dispatch Example

Send BOTH Task tool calls in a single message. Meta-judge first, implementation second:

Message with 2 tool calls:
  Tool call 1 (meta-judge):
    - description: "Meta-judge: {brief task summary}"
    - model: {meta-judge model — the user's `--model` if one was passed; otherwise same as implementation, or one tier up per Role Pairing}
    - subagent_type: "sadd:meta-judge"

  Tool call 2 (implementation):
    - description: "Implement: {brief task summary}"
    - model: {selected implementation model}
    - subagent_type: "{selected agent type}"

Wait for BOTH to return before proceeding to Phase 3.

Phase 3: Dispatch Judge Agent

After BOTH meta-judge and implementation complete, dispatch the judge agent.

CRITICAL: Provide to the judge EXACT meta-judge's evaluation specification YAML, do not skip or add anything, do not modify it in any way, do not shorten or sumaraize any text in it!

Extract from meta-judge output:

• The final evaluation specification YAML

Extract from implementation output:

• Summary section (files modified, key changes)
• Paths to files modified

3.1 Analyze the Pre-existing Changes Section

Before dispatching the judge, assess whether there are pre-existing changes in the codebase that the judge needs to be aware of. The "Pre-existing Changes" section prevents the judge from confusing prior modifications with the current implementation agent's work.

When to include:

• Previous do-and-judge task runs completed earlier in the same session
• User's manual modifications made before invoking the skill (visible from conversation context or in git)
• Changes from other tools or agents that ran before this task

When to omit:

• This is the first task with no known prior changes — omit the section entirely
• On retries within the SAME task, do NOT include the implementation agent's own previous attempt as "pre-existing changes" — those are part of the current task's iteration cycle

Content guidelines:

• Use a high-level summary: task description, list of affected files/modules, general nature of changes (created, modified, deleted)
• Do NOT include code blocks, diffs, or line-level details — keep it concise
• Label the source clearly: "Previous Task: {description}", "User modifications (before current task)", etc.
• If multiple sources of pre-existing changes exist, use separate subsections for each

CRITICAL: avoid reading full codebase or git history, just use high-level git diff/status to determine which files were changed, or use conversation context to determine if there are any pre-existing changes.

3.2 Launch Judge with prompt and specification YAML

Judge prompt template:

You are evaluating an implementation artifact against an evaluation specification produced by the meta judge.

CLAUDE_PLUGIN_ROOT=`${CLAUDE_PLUGIN_ROOT}`

## User Prompt
{Original task description from user}

{IF pre-existing changes are known, include the following section — otherwise omit entirely}

## Pre-existing Changes (Context Only)

The following changes were made BEFORE the current implementation agent started working. They are NOT part of the current task's output. Focus your evaluation on the current task's changes. Only verify pre-existing changed files/logic if they directly relate to the current task requirements.

### {Source of changes: e.g., "Previous Task: {task description}" or "User modifications (before current task)"}
{High-level summary: what was done, which files/modules were created or modified}

{END conditional section}

## Evaluation Specification

```yaml
{meta-judge's evaluation specification YAML}

Implementation Output

{Summary section from implementation agent} {Paths to files modified}

Instructions

Follow your full judge process as defined in your agent instructions!

Output

CRITICAL: You must reply with this exact structured evaluation report format in YAML at the START of your response!

CRITICAL: NEVER provide score threshold, in any format, including `threshold_pass` or anything different. Judge MUST not know what thershold for score is, in order to not be biased!!!

**Dispatch:**

Use Task tool:

• description: "Judge: {brief task summary}"
• prompt: {judge verification prompt with exact meta-judge specification YAML, and Pre-existing Changes section if applicable}
• model: {judge model — MUST equal the current implementation model, including after escalation}
• subagent_type: "sadd:judge"

### Phase 4: Parse Verdict and Iterate

Parse judge output (DO NOT read full report):

Extract from judge reply:

• VERDICT: PASS or FAIL
• SCORE: X.X/5.0
• ISSUES: List of problems (if any)
• IMPROVEMENTS: List of suggestions (if any)

**Decision logic:**

If score ≥4: → VERDICT: PASS → Report success with summary → Include IMPROVEMENTS as optional enhancements

If 3.0 ≤ score <4 and NOT STRICT_MODE: → Apply the Iteration Discretion Rule below → accepted → VERDICT: PASS (report outstanding issues) → declined → VERDICT: FAIL → go to "Check retry count" below

Otherwise (score <3.0, or score <4 with STRICT_MODE): → VERDICT: FAIL → Check retry count

If retries < 3: → Decide the retry tier per Phase 5 "Model Escalation on Retry" (bump BOTH implementation and judge, or hold) → Dispatch retry implementation agent with judge feedback → Return to Phase 3 (judge verification with same meta-judge specification)

If retries ≥ 3: → Escalate to user (see Error Handling) → Do NOT proceed without user decision

Note: `retries` counts attempts within the CURRENT cycle only. This budget **resets** on re-entry after a reported PASS (see [Phase 6 Re-entry](#phase-6-final-report)) — a later user quality complaint opens a fresh cycle of up to 3 retries even if a prior cycle already reached the limit above.

#### Iteration Discretion Rule

Your main task is to COMPLETE the task within target quality. Two failure modes are equally real:

- Burning retries and context on nitpicks so the task never completes on time → **the task is failed**; a developer is waiting on this result and may have dependent work blocked, so implementation effort MUST stay proportionate to the task's size.
- Reporting a result whose quality is genuinely too poor to be considered complete → **an even worse failure**.

Apply to every judge score: