Task Start — ground yourself before writing code

Run this at the start of any coding/engineering task. It is the compounding advantage: you begin already knowing the user's stack instead of reading cold.

Steps

1. Retrieve from the codebase. Call search_code (mcp-codebase-rag) with 2-3 phrasings of what the task touches (feature names, error messages, the modules involved). Read the returned path:line locations and snippets to learn the existing patterns. If results look thin, the repo may be unindexed — call index_repo(path) once, then search again.
2. Recall prior knowledge. For each key file/service/decision the task touches, call recall_about (mcp-knowledge-graph). Pull in any past decisions, known bugs, and relationships so you don't relitigate solved problems or repeat past mistakes.
3. Find existing conventions. Use get_symbol_context for any symbol you're about to modify, so your change matches the surrounding code's style and contracts.
4. Size the task. If it spans more than ~5 files or is architecturally non-trivial, stop and run the workflow-plan skill before coding.

Graceful degradation

If any of these tools reports unavailable (a server is down), note it briefly and proceed with normal exploration (read/grep). Never block or crash on a missing tool — these augment your workflow, they don't gate it.

Why

The win condition is persistence/compounding: retrieval-grounded context plus a queryable memory of past decisions is what a cold-start agent cannot match. Pay the small upfront retrieval cost on every task.