Runs an autonomous Creator/Validator quality loop using two CRAFT personas as independent sub-agents
Runs an autonomous Creator/Validator quality loop using two CRAFT personas as independent sub-agents. Auguste (Creator) produces content, Fernand (Validator) evaluates it against requirements, and the loop iterates until validation passes or the cycle limit is reached. Merges three former Studio recipes into a single Cowork recipe that handles the full create-validate-revise lifecycle with file-based handoffs between sub-agents.
Cowork Creator/Validator Workflow
Tags: Introduced in Beta, Creator/Validator, Quality Loop, Sub-Agent, Auguste, Fernand, Iteration, Verification, Bridge, Cowork
TL;DR
What It Does
Runs an autonomous Creator/Validator quality loop using two CRAFT personas as independent sub-agents. Auguste (Creator) produces content, Fernand (Validator) evaluates it against requirements, and the loop iterates until validation passes or the cycle limit is reached. Merges three former Studio recipes into a single Cowork recipe that handles the full create-validate-revise lifecycle with file-based handoffs between sub-agents.
How It Works
Seven steps. Step 0 validates parameters and sets up the workspace with output and cycle directories. Steps 1 and 2 build fully self-contained prompts for the Creator and Validator sub-agents (sub-agents do not inherit CLAUDE.md, so prompts must include persona instructions, formatting rules, and output specifications). Step 3 runs the main loop: the Creator sub-agent writes the deliverable and a JSON handoff summary, the Validator sub-agent reads both and produces a structured report (scoring completeness, correctness, consistency, and quality on a 0–100 scale) plus narrative feedback. The decision gate checks the verdict — PASS stops the loop, PASS_WITH_NOTES stops if no criticals, REVISION_NEEDED feeds the feedback back to Creator for the next cycle. Step 4 writes the final output. Step 5 creates an optional git checkpoint via CWK-005. Step 6 runs CWK-024 verification on the final claims. Step 7 generates an audit trail and delivery report.
What To Expect
A validated deliverable file at the configured output path, plus a cycle directory containing all intermediate artifacts (creator outputs, handoff JSONs, validator reports, feedback files) for audit purposes. The final report includes the loop status, certification level (CERTIFIED, CONDITIONAL, or UNCERTIFIED based on CWK-024 confidence), cycle count, and links to artifacts.
Best Results When You
Provide substantive requirements that include purpose, scope, constraints, and success criteria. Include specification files for the Validator to check against. Set a realistic max_cycles (3 is usually sufficient for most content). Keep auto_revise on for hands-off operation.
Typical Time
10–30 minutes depending on content complexity, number of cycles, and sub-agent processing time. Simple content with 1 cycle can complete in under 10 minutes. Complex content requiring 3 revision cycles may take 30 minutes or more.
Difficulty
Advanced — the recipe itself is autonomous, but configuring the right requirements, research files, and specification files requires understanding of the Creator/Validator pattern and what makes good validation criteria. Interpreting certification results requires familiarity with CWK-024 confidence scoring.
How To Start
This recipe orchestrates two sub-agents (Auguste as Creator, Fernand as Validator) through iterative quality cycles. It can be called directly or used by CWK-035 (CPM Phase Runner) as the execution engine for CV_WORKFLOW phases.
STEP 1Provide Content Requirements
Specify what to create, the quality requirements, and workflow preferences. The recipe handles Creator/Validator orchestration automatically.
Available parameters
-
content_type
· string · required
Type of content to create: “recipe”, “documentation”, “persona”, “cookbook”, “specification”, “template”, or “other”. Maps to formatting rules for both Creator and Validator. -
content_name
· string · required
Human-readable name for the deliverable. Used in file names, status messages, and the audit trail. -
requirements
· string · required
Detailed creation requirements including purpose, scope, constraints, and success criteria. Should be substantive — brief requirements produce brief, shallow output. -
research_files
· list · optional · default: []
Paths to reference files the Creator should read before producing content. -
specification_files
· list · optional · default: []
Paths to specs and constraints the Validator checks the Creator’s output against. -
max_cycles
· integer · optional · default: 3 · range: 1–10
Maximum Creator→Validator iterations before stopping. Clamped to range if out of bounds. -
auto_revise
· boolean · optional · default: true
When true, automatically loops on validation issues. When false, pauses for human decision at each revision point. -
confidence_threshold
· integer · optional · default: 75 · range: 50–100
Minimum CWK-024 confidence score for CERTIFIED status. Below this but above 60 is CONDITIONAL. -
output_path
· string · optional
Path for the final deliverable file. Defaults to a .cw30-output directory named after the content. -
keep_intermediate_files
· boolean · optional · default: true
Keep all cycle artifact files for audit trail. When false, only the audit trail JSON and final report are preserved. -
tier
· string · optional · default: E
Persona expertise tier for Auguste (Creator) and Fernand (Validator). B = beginner, A = advanced, E = expert. -
enable_git_checkpoint
· boolean · optional · default: true
Call CWK-005 after loop completion for crash protection. -
additional_context
· string · optional · default: ""
Extra context embedded in both Creator and Validator sub-agent prompts. -
validation_mode
· string · optional · default: comprehensive
“comprehensive” runs full CWK-024 4-gate verification. “minimal” runs structural checks only.
Invocations
Run Creator/Validator workflow
#H->AI::Directive: (Execute CWK-030 with content_type="documentation", content_name="API Guide", requirements="…")
Run with custom cycle limit
#H->AI::Directive: (Execute CWK-030 with content_type="recipe", content_name="New Recipe", requirements="…", max_cycles=5)
Run with manual review
#H->AI::Directive: (Execute CWK-030 with content_type="specification", content_name="Design Spec", requirements="…", auto_revise=false)
Usage Examples
EXAMPLE 1Creating documentation with automatic revision
User
Create API documentation for the transfer system using the Creator/Validator workflow.
Cat
Initializing CWK-030. Building Creator prompt for Auguste. Cycle 1 — Creation phase. Auguste reads reference files, produces API guide (4200 bytes). Cycle 1 — Validation phase. Fernand scores: completeness 85, correctness 90, consistency 75, quality 80. Verdict: PASS_WITH_NOTES — 2 important issues (missing error codes section, inconsistent parameter naming). No criticals. Accepting output. CWK-024 verification: CERTIFIED (avg confidence 82%). Git checkpoint created. Output saved.
EXAMPLE 2Multi-cycle revision for a recipe
User
Use the CV workflow to create a new recipe for batch file processing. Here are the requirements and spec files.
Cat
Initializing CWK-030. Cycle 1 — Auguste creates recipe draft. Fernand finds 1 critical (missing error handling), 3 important. Verdict: REVISION_NEEDED. Cycle 2 — Auguste revises with Fernand’s feedback. Fernand scores: all dimensions above 80. Verdict: PASS. CWK-024: CERTIFIED (88%). Final recipe saved after 2 cycles.
EXAMPLE 3Manual review mode
User
Run the CV workflow for the design spec but let me review each revision decision.
Cat
Initializing CWK-030 with auto_revise=false. Cycle 1 — Auguste creates spec. Fernand verdict: REVISION_NEEDED (2 critical). Cat asks: proceed with revision, accept current, or handle manually? User chooses revision. Cycle 2 — revised. Fernand verdict: PASS_WITH_NOTES. Cat asks again. User accepts.
How AI Reads This Recipe
When this recipe is triggered, the AI orchestrates a Creator/Validator quality loop with sub-agent delegation. The AI should:
- Validate and setup (Step 0). Check content_type against valid options. Verify requirements are substantive. Clamp max_cycles to 1–10. Create output directory and cycle artifact directory. Validate that all research_files and specification_files exist (warn and skip missing ones).
- Build sub-agent prompts (Steps 1–2). Creator prompt (Auguste): include role, tier, content type formatting rules, requirements, additional context, research file list, output file paths, and revision feedback (if cycle > 1). Validator prompt (Fernand): include role, tier, validation rules, requirements, spec files, creator output paths, 5 validation dimensions, issue classification, and verdict rules. Sub-agents do not inherit CLAUDE.md — prompts must be fully self-contained.
- Run main loop (Step 3). For each cycle: spawn Creator sub-agent, verify output file exists and is non-empty. Spawn Validator sub-agent, parse JSON report (fall back to text on parse error). Record cycle history. Decision gate: PASS stops loop, PASS_WITH_NOTES stops if no criticals, REVISION_NEEDED either auto-loops or pauses for human input based on auto_revise. Stop on max_cycles.
- Finalize (Steps 4–7). Write final content to output_path. Optional git checkpoint via CWK-005. Run CWK-024 verification if validation_mode is comprehensive — CERTIFIED if avg confidence meets threshold, CONDITIONAL if above 60, UNCERTIFIED below 60. Write audit trail JSON. Deliver status-specific completion report.
Error handling: if Creator produces no output or empty file, stop with ERROR. If Validator report is not valid JSON, fall back to text parsing with REVISION_NEEDED default. If files are missing, warn and skip. If git checkpoint fails, continue anyway.
When To Use This Recipe
Use this recipe when you need high-quality content that benefits from independent creation and validation by separate personas. Use it for automated iterative improvement where validation feedback drives targeted revisions. Use it when creating recipes, documentation, specifications, or other structured content where completeness and correctness matter. Use it when you want an audit trail of the creation process.
Do Not Use When
You need a quick one-off output that does not warrant multi-cycle quality review — direct creation is faster. If you only need validation without creation, use CWK-024 (Verification Gate) directly. If you need multi-phase project coordination rather than single-deliverable quality, use CWK-034 (CPM Orchestrator) with CWK-035 (Phase Runner) instead.
Recipe FAQ
Who are Auguste and Fernand in this workflow?
Auguste is the Creator persona — he produces the initial content and handles revisions. Fernand is the Validator persona — he independently evaluates the content against requirements and specs. Both run as sub-agents with self-contained prompts, so they operate independently without shared context.
What are the five validation dimensions?
Completeness (all requirements addressed), Correctness (factual and structural accuracy), Consistency (alignment with specs and patterns), Quality (professional, clear, maintainable), and Creator Requests (addressing specific validation questions the Creator flagged during creation).
What do the three verdicts mean?
PASS means no critical or important issues and all dimension scores are 80 or above. PASS_WITH_NOTES means no critical issues, 1–3 important issues, and scores above 70. REVISION_NEEDED means any critical issues, more than 3 important issues, or scores below 70.
What is the certification level?
After the loop completes, CWK-024 verification runs on the final output claims. CERTIFIED means average confidence meets or exceeds the threshold (default 75). CONDITIONAL means average is above 60 but below threshold. UNCERTIFIED means average is below 60.
Why do sub-agent prompts need to be self-contained?
Sub-agents in Cowork do not inherit CLAUDE.md or the parent session context. Each sub-agent starts fresh, so the prompt must include all persona instructions, formatting rules, file paths, and requirements needed to complete the task independently.
Can I change the cycle limit mid-workflow?
Not during an active run. Set max_cycles before starting. If you set auto_revise to false, you can stop the loop manually at any decision gate by choosing to accept the current output or handle it yourself.
Version History
v1.00a — 2026-03-02
Initial creation. Merges Studio STU-001 (Creator CVWorkflow), STU-002 (Validator CVWorkflow), and STU-003 (CV Loop) into a single Cowork recipe. Sub-agent delegation via file-based handoffs. Five validation dimensions with 3-tier issue classification. CWK-024 verification gate with confidence-based certification. Optional git checkpoints. Audit trail with cycle history.
Get this recipe with CRAFT for Claude Cowork
Cowork recipes ship bundled with CRAFT for Claude Cowork — there’s no separate download. Clone the framework once, and your AI runs every recipe automatically when invoked.
Pull anytime to stay on the latest version — free to clone, no login or email required.
Then start your session
Once CRAFT is in your project folder, open a new Cowork session and ask Claude to initialize. For example:
You
Please initialize my CRAFT session.
Claude
CRAFT session ready. Your project is loaded, your persona is active, and your recipes are available. What would you like to work on?
Not familiar with Git? Download as a ZIP
No command line needed. Just download, move, and unzip:
- Open the CRAFT framework repo on GitHub.
- Click the green Code button, then choose Download ZIP.
- Move the downloaded ZIP into your Claude Cowork project folder.
- Unzip it: double-click on Mac, or right-click → Extract All on Windows.
