Work with Hephaera day to day

Tell your agent what you want to accomplish and what should stay stable. Hephaera is designed to support that conversation with plans, relevant skills, task notes, and checks. You do not need a special phrase for an ordinary fix or feature.

A useful request includes the current problem, desired behavior, constraints, and a way to judge the result:

Players can trigger the same interaction twice while its animation is playing. Find the state owner and fix the duplicate action. Keep the animation and input bindings. Verify repeated presses, cancellation, and reopening the scene.

Ask for the kind of help you need

For investigation, make the read-only boundary clear:

Review how this feature works and explain the main risks. Do not edit files or open a PR.

For implementation, give the agent a concrete slice:

Implement the agreed settings change. Keep the public API stable, run the focused tests and build, and report the result before publication.

For an existing issue, reference it directly:

Work on #123. First import the issue into task tracking, inspect its current discussion, and plan a small change that addresses its acceptance criteria.

Hephaera keeps substantial work in task folders and links it to an issue. The agent should preserve useful evidence and unresolved questions there rather than relying on the current chat alone.

Use plans as working agreements

Check a plan for scope, ownership, dependencies, and verification. If something changes, say so while work is running:

This also ships on touch devices. Update the plan to cover touch cancellation and the smaller viewport before continuing.

You can narrow the work, ask for a different tradeoff, or stop an approach that no longer fits. A correction should update the active plan and relevant project learning. It should not disappear into an old chat message.

Roles provide a useful bias when needed. For example, ask for qa-engineer to review acceptance cases or security-engineer to inspect a sensitive boundary. Skills provide repeatable jobs; the agent checks their compatibility with the selected engine before using engine-specific APIs. See roles and skills for the catalog and resolution rules.

Choose a larger work mode deliberately

Task mode is the default for one assignment. Ask for Goal flow when you want an ordered collection of tasks with per-task PRs and post-PR QA. Goal flow can publish those task PRs after its checks, but never merges them for you.

Brainstorm mode is an explicit discovery stage for ideas that need clarification and research. Ask to enter it, develop the ideas, then explicitly exit and select an idea for planning. Exiting does not start implementation automatically.

Gauntlet mode adds an approved quality bar and independent criticism of work units. It is a separate delivery choice, with human-controlled merges, rather than a synonym for “try harder.” Read work modes and delivery before choosing it for a substantial effort.

Keep useful memory, not a chat dump

Ask the agent to retain verified conventions and hard-won fixes. Project memory lives under the host's ignored .hephaera/.cache, with relevant entries loaded by tags. Task details and debugging evidence have their own files. This is durable project context, not a promise to store every conversation.

You can inspect the available tags from the project root:

bash .hephaera/commands/query-project-context.sh --list-tags

Keep secrets out of memory, and preserve the cache when updating Hephaera. For the next step after implementation, read delivery and troubleshooting.