Behind the tools.
A workflow travels with the plugin. Your setup shapes what happens next.
Instructions, hooks, memory, conversation context, and available tools shape how an agent uses a workflow. See what comes with Fresh Curiosity Tools, what stays private, and how to adapt the setup to your own work.
What shapes a result
Installing a public package gives you its published files. It does not import my private workspace or reproduce every condition behind a demonstration.
- Plugins and skills
- The reusable methods, scripts, references and templates in your selected release.
- Instructions
- Working agreements about scope, evidence, writing, permissions and project conventions.
- Hooks
- Commands triggered at specific moments. Their code, registration, trust and actual execution are separate things.
- Conversation
- The current brief, corrections, examples, decisions and tool results available to the agent.
- Memory
- Saved preferences or earlier decisions, when your host supplies or retrieves them. My personal records are not included.
- Other agents
- Selected context and findings exchanged through bounded assignments or explicit handoffs.
Model versions, connected accounts, runtime dependencies and permissions also matter. The compatibility record distinguishes installation, representative tests and untested integrations.
Working instructions
My setup asks the agent to stay with a concrete task, use the existing source owner, check consequential facts and verify the requested outcome. Project instructions add local conventions. A visual task may need a rendered review; a writing task may need a final prose pass.
The public starter is a newly written example of those principles. It is not a copy or line-by-line redaction of my private instructions. Adapt it to your project and host before use.
Hooks and review
Published plugins already contain several hooks and optional templates. Those remain with their existing owners. A separate optional host integration shares two previously local behaviors: tracking artifact-review obligations and prompting for a scoped closeout.
A review reminder cannot judge a screenshot or prove a writing review was good. A completion mark records the caller's claim. The closeout reminder does not terminate processes or decide that a shared service belongs to this chat.
Review and configure only the hooks you want. Installing a plugin does not automatically trust its hooks, and duplicate registrations can create duplicate reminders. The integration does not activate itself.
Memory and context stay personal
In my environment, relevant preferences and earlier decisions can be supplied as context or retrieved when needed. That can reduce repetition, but an old note can also be wrong. Current files and live evidence still need checking.
My memory contents, private conversations, client records and full custom instructions are not part of these releases. You bring your own authorized context. Optional library integrations need your own data and configuration.
Even with memory off, the current conversation can influence the answer. A new chat or a shortened history may lack a crucial decision. A concise handoff should name the goal, accepted constraints, source and unresolved question.
Between agents and chats
A lead can assign a worker a narrow question and selected context, then inspect the returned evidence. Workers may receive different amounts of history. Separate chats communicate through explicit tools or maintained handoffs; they do not automatically share everything.
In this workflow, a useful lesson changes future behavior when it is deliberately retained or used to update a maintained method. Exchanging messages does not itself retrain the model or update every chat. Several agents agreeing can still mean they shared the same incomplete evidence.
Each assignment needs an owner, inputs, expected output and stop condition. Sending material to another provider is a separate data-sharing decision.
Try your own setup
- Start with one package. Record the release, host and model. Follow its installation guide.
- Use a fictional task with explicit facts. A fresh chat may still receive global instructions or memory. Inspect your settings.
- Check what actually ran. Identify the selected skill, tools and hook state. Review the output against your acceptance criteria.
- Add one influence at a time. Compare the same task before adding optional hooks, private sources or extra agents.
For help, share the version, sanitized task, expected result and observed result. Keep private instructions, memories, credentials and raw account logs out of a public issue.