dont-know-how
The user faces a task they cannot start: an unfamiliar system, protocol, library, or integration. Your deliverable is direction, given in the conversation: the real evidenced approaches the sources support, with a recommendation. Implementation is out of scope.
One rule governs every step: evidence, never guesswork. Every claim carries a locator: a path:line, a URL, a document section, or a user statement. A fact you could not establish is reported as unknown, not filled in.
1. Scan the repo
Extract the keywords from the prompt: the named system, its synonyms, the protocols and standards it implies, and the vendor or product names. "101 AD server" expands to Active Directory, LDAP, LDAPS, Kerberos, domain, bind, and the names of libraries the project's language uses for them.
Search with the strongest tool the environment offers: a code knowledge graph when the project has one, otherwise whatever file and content search is available. Cover code, configuration, dependency manifests, and documentation. Look for what the project already owns that touches the task: existing modules, config entries, installed packages, README sections, prior notes.
Done when every asset found is listed with its locator, or the absence of anything relevant is stated plainly.
2. Ask what the user holds
Ask only for what search cannot reach, and only when the answer could change a candidate direction, its feasibility, or a material risk: documents from the counterpart, sample code, test environments and credential availability, contacts, and constraints. If no remaining question clears that gate, skip this step and continue to step 3.
Build the round from references/questions.md: pick the template matching the task shape, drop questions the scan already answered or the directions do not depend on, and word each remaining question in the task's own terms. For credentials, ask only whether suitable access exists, how it is obtained, and who manages it. Keep passwords, API keys, private keys, tokens, and certificate secrets out of the conversation.
Then wait for the answer.