Search Before Creation

Before implementing new functionality or reading literature, the system encourages discovery of existing project knowledge.

Typical order

1. Search local code intelligence (codio registry)
2. Inspect prior notes and documentation
3. Check external libraries
4. Consult literature
5. Implement new work

This reduces duplicated effort and improves reuse.

From discovery to decision

Discovery results feed into explicit engineering decisions:

• Reuse existing internal implementation
• Wrap an external library with a project-specific interface
• Depend on an external library directly
• Implement new code

Projio aims to turn discovery into actionable engineering decisions rather than leaving agents to guess.

Agent workflows

The MCP tool surface supports this workflow directly:

1. codio_list / codio_discover — search for existing implementations
2. codio_get — inspect curated library notes
3. rag_query — search code and documentation corpora
4. note_search — check prior design decisions
5. citekey_resolve / paper_context — consult literature

Parallel agents can study multiple libraries simultaneously. Results feed back into curated library notes for future reference.

From tools to workflows

Having the right tools is necessary but not sufficient. The key bottleneck for autonomous agent productivity is workflow awareness — knowing when to use which tools in what sequence.

Projio addresses this at three levels:

1. Workflow prompts

Composable, phase-specific instruction documents at docs/prompts/workflows/ that agents can read to self-direct:

Prompt	Phase	Purpose
session-bootstrap.md	Any	Context gathering at session start
explore-idea.md	Exploration	Capture idea, discover code, search literature, decide
implement-feature.md	Exploration	Implement, test, notebook demo
integrate-pipeline.md	Production	Snakemake integration, registry, docs
validate-and-deploy.md	Production	Pre-flight, validation, deployment

These chain naturally: each prompt's output becomes the next prompt's input.

2. Cross-package skills (planned)

Python functions at src/projio/skills/ that compose tools from multiple packages into a single call. For example, explore_idea("phase gradient detection") internally calls codio_discover + rag_query + note_create and returns a structured report.

3. Phase-aware routing (planned)

The agent_instructions() MCP tool will detect the current workflow phase from recent notes and active tasks, then suggest the appropriate workflow prompt — so the agent knows where it is in the routine without the user having to explain.

See the Grand Routine tutorial for the full end-to-end workflow.