Workflow Preflight

Run this preflight before starting substantive work. Its purpose is to improve the main task by selecting the right capabilities, not to load every available skill or delay straightforward work with unnecessary research.

Select skills before acting

1. Identify the outcome, target environment, data sensitivity, external systems, and evidence needed to call the task complete. Stop here for a direct answer, calculation, or isolated rewrite: no specialized skill is needed.
2. Make a provisional shortlist from the available-skills metadata. Prefer the published inventory; for a workspace, also check project roots such as .agents/skills/, .claude/skills/, .opencode/skills/, and .hermes/skills/. Limit the shortlist to the few capabilities that plausibly cover a material part of the task. Never select a route whose skill or required tool is unavailable in the current runtime.
3. Read the full instructions only for that shortlist. Select the smallest compatible set based on coverage, sequencing, available tools and access, and the cost of extra setup. A group of skills must work together for this request; individually relevant skills can still be redundant or conflict.
4. State the route in one short line, such as Preflight: graphify + product-quality-loop — architecture context and verified implementation. Then begin the main task in the same turn.
5. Use agent-workflow-pipeline when the task combines repository context, public research, implementation, optional model routing, and verification. Treat it as the coordinator; do not separately load a component skill when the pipeline already owns that phase.
6. Only if no selected skill covers a material capability, read skill-discovery.md. Recommend a verified candidate, but do not install it, connect a service, or add credentials without the user's explicit authorization.

Use evaluation.md when changing this skill or checking its routing behavior. The test cases separate trigger accuracy from route quality, so a broad description does not turn into indiscriminate skill loading.

Open the working context

Every task that passes step 1 runs on the working-context-notebook flow. It is the default working flow, not a route to select, so it does not count toward the smallest skill set and is not named in the route line.

• Open the notebook as soon as the route is stated. Put the outcome in the goal and the completion evidence from step 1 in the tracker.
• While the selected skills run, log long raw output, keep exact values as sourced facts, and accept an edit at each phase or gate.
• Read the accepted revision back after any context compaction and before the final answer.

Skip it only where step 1 stops: a direct answer, a calculation, or an isolated rewrite.

Common local routes

• Repository architecture, file relationships, or codebase questions: graphify.
• Current public facts or structured source research: scrapegraphai-research.
• Platform content (YouTube, GitHub, Reddit, Twitter/X, Bilibili, RSS), or installing or updating Agent Reach: scrapegraphai-research (see scrapegraphai-research's references/agent-reach.md).
• User-authorized feature work, fixes, or release preparation: product-quality-loop.
• Coder workspace or TrueForge handoff: remote-agent-graphify.
• Cross-agent work that needs several of the routes above: agent-workflow-pipeline.
• Building or substantially extending a product end to end (research, questions, jury-reviewed tasks, testing, integration, graph, memory): engineering-workflow.
• Profiling or prioritizing clients from a CRM export: client-profiling.
• Browser interaction or web-app testing: agent-browser.
• A repeatable end-to-end test suite for a web or mobile app: npx e2e init through product-quality-loop (see product-quality-loop's references/e2e-testing.md). It installs packages, a skill, and an MCP entry, so treat it as an install that needs the user's go-ahead unless they asked for it.
• Frontend design or UX work: impeccable.
• Security assessment: application-security-testing.

These are starting points. Use the task wording and each skill description to make the final selection, including other installed skills that fit better.

Keep the main task moving

Preflight is complete once the selected skill set is clear. Continue into the main task immediately. When an online candidate would materially improve the result, present its source, reputation signals, scope, and installation command as a recommendation, then wait for the user to ask for installation.