Skip to content

Latest commit

 

History

History
71 lines (50 loc) · 3.2 KB

File metadata and controls

71 lines (50 loc) · 3.2 KB

Instructions

This directory contains self-contained agent prompts used by the autonomous recipe generation system.

Each file is a complete, standalone prompt that can be executed by an AI agent (Claude) either automatically via a GitHub Actions workflow or manually from your local machine.

Available Instructions

File Triggered By What It Does
pm.md Every hour at :07 Scans recipes/ for coverage gaps and creates queue:generate issues for missing recipes. Also detects new Deepgram SDK repos not yet tracked.
engineer.md Every hour at :27, on issue events Routes open type:queue issues. Executes generate-examples.md for action:generate issues; surfaces action:new-sdk issues for human review.
generate-examples.md Triggered by queue issues Generates runnable recipe files (example.*, example_test.*, README.md) for a language/product combination. Opens a PR and closes the queue issue.
lead-coverage.md On PR merge to main Rebuilds COVERAGE.md and per-language/per-product README files to reflect the current state of recipes/.
lead-reconcile.md Daily at 11:45 UTC Verifies the root README.md count table is accurate. Creates queue:fix issues for sample directories missing required files.

How to Run Locally

Any instruction can be run locally using the Claude CLI:

claude --model claude-opus-4-6 -p "$(cat instructions/{name}.md)"

Examples:

# Check for coverage gaps and create queue issues
claude --model claude-opus-4-6 -p "$(cat instructions/pm.md)"

# Process the oldest pending queue issue
claude --model claude-opus-4-6 -p "$(cat instructions/engineer.md)"

# Generate examples from a queue issue
claude --model claude-opus-4-6 -p "$(cat instructions/generate-examples.md)"

# Rebuild coverage documentation after merging samples
claude --model claude-opus-4-6 -p "$(cat instructions/lead-coverage.md)"

# Reconcile the index and check for missing files
claude --model claude-opus-4-6 -p "$(cat instructions/lead-reconcile.md)"

Design Principles

  • Instructions are idempotent — running the same instruction twice should not create duplicates.
  • Instructions check existing state before taking action (no blind writes).
  • Instructions never commit directly to main.
  • Instructions skip failing items with a warning rather than aborting entirely.

Cron-as-safety-net pattern

GitHub blocks event propagation between workflows when GITHUB_TOKEN is the actor. A workflow-created issue does not fire on: issues. A workflow-merged PR does not fire on: pull_request.closed. Any autonomous-workflow → event → other-workflow chain is silently broken.

Every downstream workflow has a cron fallback for this reason:

pm (cron :07)
  └─ creates queue:generate issues
       └─ engineer (cron :27) ← cron fires even when issues were created by a workflow
            └─ merges recipe PRs
                 └─ lead-coverage (cron every 6h) ← cron fires even when PR was merged by workflow

lead-reconcile (daily) is the final safety net — catches any drift the other workflows missed.

When adding new downstream workflows, always include a cron trigger.