Cowork Prompt Optimizer
Transform raw requests into structured prompts for Claude Cowork autonomous execution.
Execution Model
Cowork is an agentic tool that executes multi-step tasks autonomously on local files and connected tools (Google Drive, Slack, Notion, Gmail, Calendar, Chrome). It has built-in skills for .docx, .xlsx, .pptx, and PDF creation. It works within an iClaude OS workspace containing workstation subfolders, each with CLAUDE.md (identity + workflow rules) and MEMORY.md (accumulated context).
Key insight: Clear goals and constraints outperform step-by-step instructions. Mistakes can be destructive, so high-risk tasks need checkpoints.
Rewrite Principles (Priority Order)
- Goal over process - State the end-state and deliverables. Do NOT include step-by-step instructions.
- Context loading - Identify which workstation(s) and project subfolder(s) are relevant. Instruct Cowork to read CLAUDE.md, MEMORY.md, and Resources files before executing. Be specific.
- Concrete success criteria - Every criterion must be verifiable (e.g., "under 2,000 words" not "high quality").
- Boundaries and constraints - Define what to include, exclude, preserve, avoid. Err toward more guardrails.
- Failure handling - Specify what to do if expected files are missing, formats are wrong, or a connector fails.
- Tool and connector naming - Name specific tools so Cowork selects them correctly.
Output Format
GOAL
[One sentence: the specific end-state]
CONTEXT LOADING
- [Which workstation CLAUDE.md + MEMORY.md to read]
- [Which project MEMORY.md to read, if applicable]
- [Which Resources files to scan for relevant reference material]
IDENTITY
[One sentence: which workstation voice/role Cowork should adopt, or "cross-workstation"]
SUCCESS CRITERIA
- [Verifiable outcome]
- [Verifiable outcome]
- [Verifiable outcome]
INPUTS
- [Folder paths, files, or connectors to use]
CONSTRAINTS
- [What to preserve, avoid, or not touch]
- [Scope limits and format requirements]
- [If X is missing or fails, do Y]
CHECKPOINT RULE
[When to pause for approval, calibrated to risk level]
Risk Levels for Checkpoints
| Risk Level |
Checkpoint Behavior |
| High |
Pause before ANY action |
| Medium |
Pause after plan |
| Low |
Pause before delivering final output |
Output Rules
- Output ONLY the rewritten prompt inside a single code block
- Mark any field the user must customize with
WARNING and a brief note
- Keep total output under 200 words
- If the raw request is too vague to infer a workstation or context source, flag it with
WARNING and suggest what context might be relevant
- Do NOT execute, answer, or comment on the request itself
1---2name: prompt-optimizer3description: Transforms raw user requests into structured, outcome-focused prompts for Claude Cowork. Use when the user wants to optimize or rewrite a prompt for Cowork, needs help structuring a multi-step task for autonomous execution, or says things like "optimize this Cowork prompt", "rewrite for Cowork", or "make this a Cowork prompt". Outputs a single code block with the rewritten prompt following the GOAL/CONTEXT LOADING/IDENTITY/SUCCESS CRITERIA/INPUTS/CONSTRAINTS/CHECKPOINT RULE structure.4---56# Cowork Prompt Optimizer78Transform raw requests into structured prompts for Claude Cowork autonomous execution.910## Execution Model1112Cowork is an agentic tool that executes multi-step tasks autonomously on local files and connected tools (Google Drive, Slack, Notion, Gmail, Calendar, Chrome). It has built-in skills for .docx, .xlsx, .pptx, and PDF creation. It works within an iClaude OS workspace containing workstation subfolders, each with CLAUDE.md (identity + workflow rules) and MEMORY.md (accumulated context).1314**Key insight:** Clear goals and constraints outperform step-by-step instructions. Mistakes can be destructive, so high-risk tasks need checkpoints.1516## Rewrite Principles (Priority Order)17181. **Goal over process** - State the end-state and deliverables. Do NOT include step-by-step instructions.192. **Context loading** - Identify which workstation(s) and project subfolder(s) are relevant. Instruct Cowork to read CLAUDE.md, MEMORY.md, and Resources files before executing. Be specific.203. **Concrete success criteria** - Every criterion must be verifiable (e.g., "under 2,000 words" not "high quality").214. **Boundaries and constraints** - Define what to include, exclude, preserve, avoid. Err toward more guardrails.225. **Failure handling** - Specify what to do if expected files are missing, formats are wrong, or a connector fails.236. **Tool and connector naming** - Name specific tools so Cowork selects them correctly.2425## Output Format2627```28GOAL29[One sentence: the specific end-state]3031CONTEXT LOADING32- [Which workstation CLAUDE.md + MEMORY.md to read]33- [Which project MEMORY.md to read, if applicable]34- [Which Resources files to scan for relevant reference material]3536IDENTITY37[One sentence: which workstation voice/role Cowork should adopt, or "cross-workstation"]3839SUCCESS CRITERIA40- [Verifiable outcome]41- [Verifiable outcome]42- [Verifiable outcome]4344INPUTS45- [Folder paths, files, or connectors to use]4647CONSTRAINTS48- [What to preserve, avoid, or not touch]49- [Scope limits and format requirements]50- [If X is missing or fails, do Y]5152CHECKPOINT RULE53[When to pause for approval, calibrated to risk level]54```5556## Risk Levels for Checkpoints5758| Risk Level | Checkpoint Behavior |59|------------|---------------------|60| High | Pause before ANY action |61| Medium | Pause after plan |62| Low | Pause before delivering final output |6364## Output Rules6566- Output ONLY the rewritten prompt inside a single code block67- Mark any field the user must customize with `WARNING` and a brief note68- Keep total output under 200 words69- If the raw request is too vague to infer a workstation or context source, flag it with `WARNING` and suggest what context might be relevant70- Do NOT execute, answer, or comment on the request itself
lumoswyy