/ratel-assessment — front-door audit of a partner agent codebase

The first conversation with a partner startup is rarely "wire up Ratel." It is "show us you understand our agent." This skill is the front door: it reads the codebase, runs every dimension in the assessment catalog, and produces a markdown report the partner could (and would) share internally — whether or not they end up using Ratel.

The deliverable is a polished, evidence-led report at <repo>/.ratel/ratel-assessment-<YYYY-MM-DD>.md, rendered alongside a branded, scored HTML version at <repo>/.ratel/ratel-assessment-<YYYY-MM-DD>.html (gauge, radar, per-dimension score bars). The report carries credibility because every finding cites a file path, a line range, or a concrete count. It carries narrative because the Ratel-relevant findings cluster into a "Where Ratel fits" section that ties them to specific shipped features and roadmap versions. It carries momentum because it ends with conditional pointers to the right follow-up skill from the suite (/ratel-observability-assessment, /ratel-integrate, /ratel-decompose-prompt, /ratel-tune-definitions).

This skill complements the rest of the suite:

Runs before the other four. It is the only one safe to fire on a cold repo with no setup from the partner.
Routes into the rest. The "Recommended next steps" section names which follow-up is appropriate given what was found.
Does not duplicate the catalogs of the others. Ratel angles point to ../ratel-observability-assessment/references/ratel-value-map.md; naming references point to ../ratel-observability-assessment/references/semantic-conventions.md.

Philosophy

Three rules. Break any of them and the report becomes a sales document, which the partner will see through immediately.

Every finding must cite evidence. A file path with a line range, a count, or a quote from the codebase. Vague claims ("consider improving observability") earn nothing. Specific claims ("47 tools registered across src/agents/*.ts, none with description longer than 12 words") earn the next meeting. If you can't cite, drop the finding.
Honest skip beats fabrication. If the codebase is not an agent, say so and stop — do not invent findings. If a dimension has nothing wrong with it, mark it Strong and move on; do not stretch to find a flaw. Partners notice the absence of fluff faster than they notice its presence.
Ratel is a finding category, not a conclusion. Ratel angles attach to findings that genuinely match Ratel's surface (tool sprawl, retrieval, decomposition, observability). They do not get bolted onto every finding to inflate the "Where Ratel fits" section. If only two findings legitimately map to Ratel, the section has two entries — not eight padded ones.