spb/llm-api-toolkits

1# Changelog23## 1.0.1 — 2026-11-1045- Added `examples/`: two screenshots of real sessions (Claude Code and Codex) using version 1.0.0 installed from the registry, declared in the manifest so the registry page shows them. No change to the code.67## 1.0.0 — 2026-11-1089First release.1011- Model and rule tables generated from live probes of 15 models (46 request shapes each).12- Offline CLI: `models`, `lint`, `fix`, `migrate`, `cost`, `compare`, `scaffold`, `guide`.13- 14 scaffold templates (7 approaches × TypeScript and Python), each executed against the live API.1415### Measured effect1617An author model was asked to write Messages requests for 16 tasks, each targeting a specific model,18with and without this skill in its system prompt. Every request was sent to the live API; a request19counts when it passes validation (no HTTP 400).2021| Author model | Without skill | With skill |22|---|---|---|23| `claude-haiku-5-5` | 8 / 16 | 16 / 16 |24| `claude-sonnet-5-5` | 7 / 16 | 16 / 16 |25| `claude-opus-5-5` | 8 / 16 | 16 / 16 |2627Read these numbers with their limits:2829- The 16 tasks were written by the skill's author and lean towards known pitfalls (sampling30  parameters, thinking modes, forced tool choice, prefill). They are not a random sample of work.31- The skill was revised twice after earlier runs on these same tasks: a first run scored 15/1632  (Haiku) and 14/16 (Sonnet). The failures exposed a missing beta-flag table and an unclear33  placement rule for mid-conversation system messages; both were fixed, and the linter gained a34  rule. The final figures are therefore measured on tasks the skill was tuned against.35- One run per condition; model output varies between runs.36- "Passes validation" is not "does the task well". Output quality was not graded.37- For the fast-mode task the account had no fast-mode quota: the API answered 429, which is counted38  as a valid request.3940The task list, harness and raw results are in the source repository (`anthropic/skill-tests/`).41