Zorya Context

Context guide

Use Context when saved knowledge and specialist perspectives can help with a task. Start with a small, checkable result. Add roles only when they help with this task.

Setup Games Workflow video

Quick start

Connect Zorya Context using the setup instructions and confirm that your client shows its tools.
In a new request, explicitly invoke the installed Context plugin and ask it to use all applicable saved knowledge, skills, and project rules.
Describe the task, audience, constraints, and checkable result. Name relevant specialists or ask the model to choose from those available.
Start with a short recipe below. Review the result and ask for a specific revision if it misses the goal.
Start with saved knowledge. Separate consent is required before each new research task or collection of new information.

Prompt recipes

The site's games are separate interactive demonstrations. The recipes below are copyable prompts for your own task.

1. A small new website

Try Context on a concrete task: plan a page for a neighborhood tool library using relevant saved knowledge. The example is complete and ready to copy.

1. A small new website

Invoke the installed Zorya Context plugin. Read all applicable skills and project rules. Start with relevant knowledge already saved in Context. If new sources or research are needed, ask for separate consent for that specific, bounded collection first. If no relevant saved knowledge is available, say so; do not invent it.

Prepare a text draft for a small one-page website for a neighborhood tool library. The audience is local residents who want to learn how to borrow a tool. The page goal is to explain the borrowing rules and help visitors contact a volunteer librarian. Constraints: do not add sign-up, payment, or tool availability details that were not provided. Prepare a headline, short explanation, three sections, and one primary contact button. Explain which relevant saved knowledge informed the structure. Check that the page serves its goal and includes no invented terms.

You can copy this example as written. For your own task, replace the audience, goal, constraints, and result.

2. Improve an existing page or code

Ask Context to identify one small, useful change for a real page and carry out the assigned work within the project rules.

2. Improve an existing page or code

Invoke the installed Zorya Context plugin. Read all applicable skills and project rules. Start with relevant knowledge already saved in Context; start by reusing relevant material. Before each new research or collection from new sources, request separate consent for that specific collection.

Inspect [page, path, or code area] and propose one small improvement for the user task: [goal]. Account for this audience: [audience] and these constraints: [constraints]. Use existing components if present, but do not assume they are required. If your AI client or app supports a separate consultant, ask them to assess the idea; if not, state that limit and do not present your own extra pass as an independent review. Make the assigned reversible change within this task and the project rules. At the end, report the files changed, what you checked, and what remains unchecked.

3. Iterate on a design

Ask the agent to explain why it chose this version, suggest a specific improvement, make it, and check the result.

3. Iterate on a design

Invoke the installed Zorya Context plugin, read the applicable skills and project rules, and use relevant knowledge already saved in Context. Before each new research or collection from new sources, request separate consent for that specific collection.

Review [the design version in this conversation or the attached image]. Explain what intent you think it expresses and why those choices may have been made. If the author is unknown, do not assign them motives: distinguish what is visible from your assumptions. Compare the visible design with the original goal: [describe the goal]. Suggest one specific improvement, make the assigned revision, and check it against the goal. If your AI client, app, or environment supports a separate designer or independent evaluator, ask for a separate review; otherwise, say so and do not present your own extra pass as an independent review. At the end, report what you actually checked and what remains unknown.

How to improve a first result

Ask the agent to explain the intent behind the current version and why it chose those options. If the author is unknown, have it describe what is visible and separate that from assumptions. Then compare the version with the original goal, give specific feedback, propose one improvement, and check the revision. State what remains unknown or unchecked.

Roles in a workflow

A role describes a responsibility, not a guaranteed feature. Available capabilities depend on your AI client, app, or the environment where the agent runs.

Orchestrator: Clarifies the goal, breaks work into steps, directs questions to specialists, combines their input, and keeps the work within scope.
Specialist consultants: Answer a focused question in their field using saved knowledge. For example, an editor checks clarity while a UX specialist checks the user journey.
Collectors and evaluators: Select and compare relevant knowledge already saved in the Context library. Gathering new sources or doing research requires separate consent for each specific, bounded collection.
Executor: Carries out the agreed work within clearly stated files and boundaries. For software, that means changing code; for a writing task, it means preparing the final text.
Independent quality check: Reviews the result separately from the executor and reports concrete evidence: what was checked, what was found, and what remains unknown.

Check what your AI client, app, or environment actually supports. Assign a separate model or launch a separate consultant only when that is available and project rules allow it. If a separate role is unavailable, state the limit: another pass by the same model is not an independent consultation or quality check.

Optional: an advanced workflow

For a larger task, you can time-box saved-knowledge gathering to one to three hours. This limit applies to context gathering, not the full implementation.

More context and additional roles can increase model time and cost. Try a short example first; expand the workflow only when it helps.

Optional: an advanced workflow

Invoke the installed Zorya Context plugin. Read and follow all applicable skills, project rules, and the actual minimum requirements of the chosen workflow. Start with relevant knowledge already saved in Context and follow Context rules for preparing that volume, including consent requirements for an extended reuse_only step. Before each new research task or collection of new sources, agree with me on the specific gap, benefit, sources, and time limit; each such collection requires separate consent.

Assign responsibilities to an orchestrator, relevant specialist consultants, a knowledge collector and evaluator, an executor, and an independent quality check only when the AI client, app, or environment and project rules actually support those roles. If a role is unavailable, state that limit; the executor reviewing its own work in sequence is not an independent consultant or QA.

Do not treat the preference for saved knowledge as blanket consent for extended preparation; follow applicable Context rules and ask separately if they require it. Optional Context target: 100,000 unique relevant knowledge tokens for each role separately, including the executor. Count only complete bodies actually read by that role; exclude duplicates and irrelevant bodies. Time-box knowledge gathering itself to one to three hours. When that time ends, report the time and actual volume for each role, including any shortfall, then continue the authorized implementation with the gathered material if the workflow's mandatory minimums and dependencies are satisfied. If a required minimum or separate role is missing, do not claim that step passed and state which dependency prevents continuation.

Optional assignment of available models: Orchestrator → [available model]; consultants → [available model]; collector/evaluator → [available model]; executor → [available model]; independent QA → [available model]. Fill this in only if your AI client, app, or environment supports model selection and project rules allow it. The same model may serve multiple separate roles or agents if supported; QA independence comes from a separate role and review, not a different model. This assignment does not change MCP tools or guarantee separate agents can be launched.

Task: [goal, audience, constraints, affected files, and acceptance criteria].

Example model assignment

Only if your AI client, app, or environment supports model selection and project rules allow it: Orchestrator → [available model]; consultants → [available model]; collector/evaluator → [available model]; executor → [available model]; independent QA → [available model]. The same model may be used in multiple separate roles/agents if supported. QA independence comes from a separate review, not the model.