project-mentor

Project mentor

This skill explains OpenReel Studio project rules and points to the right local guide when the model needs more than workflow/prompt skills.

Default production is node-first: the model works on one visible canvas with text, image, video, and audio nodes. It does not create a separate planning object or maintain separate canvas/panel state before work appears.

Topics

• overview: repo layout, commands, and current architecture.
• agent_loop: Agent loop, core tools, permission policy, trace, and compaction.
• production_audit_guide: final delivery and consistency review.
• node_repair_guide: failed node repair, rerun, and dependency recovery.
• slash_commands: deterministic command surface.
• debugging: trace, SSE, messages, tool results, and artifacts.
• prompt_compaction: prompt/cache budget and where rules should live.

Current Rules

• Ordinary image/video work uses the automatically injected Skill catalog, chooses the minimal description match, resolves its exact package with skills.list, and reads the complete SKILL.md with skills.read. It then creates lightweight tasks for multi-step or media-generation work, then creates or updates text, image, video, and audio nodes directly on the canvas.
• Main Agent plans the node graph and dependency order. Each node is an independent task. Script, character image, scene image, shot grid image, and final video prompt are produced by node.run with one module prompt skill at a time.
• Reusable graph workflows are executed through tool.search, tool.describe, and tool.execute for deferred workflow.run_step, workflow.run_next, or workflow.run_all; the workflow runner calls node.run internally for visible product nodes.
• Canvas state is the creative truth source visible to the model. Drafts, grouping, method choice, review notes, and assumptions are node fields or text nodes, not a separate planning object.
• Dependencies are expressed with parent_node_id and fields.references; backend-created edges appear automatically. Use {ref, role} when needed: visual_reference for generation reference, source_image when an image node directly adopts an existing image as output.
• Project-local node numbers such as #0 / 0 resolve directly with node.get(node_id); titles or unclear references use node.list(query|regex).
• Read the smallest sufficient evidence for each decision: aggregate state, then indexes, details, and only the required content page.
• Use interaction.request_input(questions=[...]) for blocking missing facts, then wait; ask up to 6 concise questions.
• An explicit destructive request calls its matching tool once at the intended scope; that first call creates structured confirmation and ends the turn.
• Generated media remains in node output and local project storage by default. Save to the asset library only when the user explicitly asks.
• Natural-language tasks enter the Agent loop. Backend preprocessing may clean input and stale state, but it must not decide business actions for the model.
• Summaries, rewrites, formatting, analysis, and plans return in chat unless the latest user message explicitly asks to save or change canvas content. An explicitly saved long text uses a placeholder text node with fields.generation={instruction,source_message_count}, followed by node.run. The count covers the source message and current save request (usually 2 for a follow-up request); the runner captures those message ids and atomically saves only a complete result. A successful run is final and does not need node.get verification. Text bodies are returned once in content_page, with an 8,000-character default window. Continue with content_offset=content_page.next_offset and a bounded content_limit; content_limit=0 returns body metadata only.
• Tool errors are observations. Read error_kind, hint, and suggested_next; repair the specific node or field before retrying.
• Every tool result crosses one typed model-context compiler. JSON, documents, collections, and multimodal parts have executable per-tool policies and a global 10,000-token hard ceiling. Large raw results are retained only as project-scoped diagnostic artifacts; model context and SSE receive bounded projections plus an opaque artifact_ref. Only resumable page content gets the document-sized string window; unrelated nested strings keep the normal per-field ceiling.
• Long text readers such as node.get, file readers, text assets, and workflow spec/template readers expose bounded content pages with next_offset. skills.read uses the opaque next_cursor returned by its preceding page; continue until it becomes null.
• User skills use skills/<skill-name>/SKILL.md. Standard frontmatter supplies name and description; the kebab-case directory name matches name. Supporting files stay inside the same package under references/, scripts/, assets/, templates/, or agents/openai.yaml. The runtime prompt carries only bounded metadata. Current-turn $SkillName, linked skill:// / SKILL.md mentions, and structured kind=skill inputs resolve before the model call and inject each selected package as a bounded <skill> block. Multiple explicit selections are all injected, and they do not carry into later turns. Explicit selection remains available when agents/openai.yaml disables implicit invocation. Description matches and plain-text names use the model-visible catalog. Orchestrator resources use exact handles from skills.list; skills.read follows next_cursor through the full SKILL.md and resolves resources inside the same package.
• Collection readers use bounded pages. project.get_state returns runtime state plus canvas counts rather than every node and edge; use node.list and node.get for details. Most collections expose offset/next_offset or a nested page; skills.list uses its opaque next_cursor.
• System prompt stays short. Detailed workflow, examples, and debugging advice live in skills, docs, tests, validators, and permission policy.