ai-config-builder

# AI config builder Use this skill when creating or changing reusable AI configuration. ## Modes Choose the mode that matches the project's current state. ### Bootstrap Use when a new or empty project needs the minimum AI configuration required to begin working safely. Select only immediately useful configuration. Do not use Bootstrap to discover the product itself. Use `grill-me` when product or technical decisions are unresolved. Do not recommend stack-specific or workflow-specific configuration before decisions justify it. ### Adopt Use when the project is defined or already exists. Inspect the actual project and classify relevant configuration as: - ADD - KEEP EXISTING - REPLACE - SKIP Project-specific behavior takes priority over reusable defaults. ### Build Use when creating or extending a skill, agent, rule, template, or Project Instruction. Inspect: - `README.md`; - related existing configuration; - the target project's context. Then choose: - CREATE - EXTEND EXISTING - DON'T CREATE Do not create a new artifact when an existing one already owns the responsibility. ### Create a ChatGPT or Claude Project Use when setting up a long-running workspace with multiple related chats. This is a Project Instruction task, not a request to copy a whole preset unchanged. 1. Select the closest existing preset from `README.md` and `WORKFLOWS.md`; use `project-base` only when no specialized preset fits. 2. Identify the project's purpose, scope, audience, and sources of truth from supplied files and explicit decisions. Record missing context rather than filling gaps. 3. Separate durable workspace behavior from reference documents. Keep instructions focused on how chats should work; keep detailed facts, research, and evolving project material in referenced files. 4. Preserve decisions and rationale in a durable project source so they carry across chats. Distinguish decisions from findings, hypotheses, and open questions. 5. Remove generic preset instructions that do not apply. Avoid repeating context already present in project files or another authoritative instruction. 6. Add only project-specific customization needed to resolve actual requirements or constraints. Keep project-specific behavior authoritative over reusable defaults. 7. Return concise final Project Instructions, the chosen preset, the sources used, and any unresolved questions. ## Identify the artifact type Choose: - Skill — procedure for a specific task. - Agent — specialized role or isolated responsibility. - Rule — stable constraint. - Template — reusable artifact structure. - Project Instruction — long-lived behavior for a multi-chat project. ## Discovery Ask only the questions needed to define: - purpose; - trigger or use case; - required context; - expected behavior; - boundaries; - output; - completion criteria; - tools or related skills; - how it will be tested. For agents also determine: - what responsibility is isolated; - what the agent may modify; - what requires approval; - when it must stop; - what it reports back. ## Writing Keep configuration compact. Prefer: - clear responsibilities; - behavior over explanation; - references to canonical sources; - progressive disclosure. Avoid: - duplicated rules; - giant instruction files; - speculative capabilities; - implementation details that belong in project code; - copying documentation into prompts. Use `writing-for-agents` when useful. ## Adoption New configuration starts as `testing`. Do not mark it accepted or default without explicit user confirmation and real usage evidence. ## Return Match the output to the selected mode. ### Bootstrap Provide: - minimum starting configuration; - why each item is needed now; - decisions intentionally deferred until the project is clearer; - next discovery step. ### Adopt For each relevant item provide: - ADD / KEEP EXISTING / REPLACE / SKIP; - why it applies; - overlap with existing project behavior; - where it should live. ### Build Provide: - CREATE / EXTEND EXISTING / DON'T CREATE; - artifact type and name; - overlap with existing configuration; - proposed file location; - finished content when creation is approved; - one realistic test scenario.