# Claude Opus 5.5 · v1 run · 28 September 2026

Five briefs from [`../prompts.md`](../prompts.md), built twice: first with no optional skills (`claude-opus-5-5/`), then after reading the pinned official skill (`claude-opus-5-5/skills/`). Same filenames in both folders.

| | |
| --- | --- |
| Model | `claude-opus-5-5` (Claude Opus 5.5), Claude Code desktop app |
| Reasoning | Medium, as requested by the user. The session cannot read or change its own effort setting, so this is the requested value, not a verified one. |
| Skill | [`benchmark/skills/anthropic-frontend-design-v1/SKILL.md`](../benchmark/skills/anthropic-frontend-design-v1/SKILL.md), SHA-256 `d91970639e9f5c37682ac7ab60094d35f1c7c1f38d731bd56396563aee10c1d3` |
| No-skills stage | 18:39 to 18:51 UTC |
| Skill stage | 18:51 to 19:03 UTC, including reading the skill and writing [`skills/DESIGN.md`](skills/DESIGN.md) |
| Checks and repairs | 19:03 to about 19:15 UTC |
| Token usage | Unavailable |

## Session design and confounders

- One session for everything, matching the earlier v1 Sol, Luna and Astra runs: all five no-skills pages first, then the skill, then all five skill pages. This is a sequential comparison, not an isolated skills-on/off experiment. The skill stage had the no-skills work in context.
- The same session first cleaned up this repository. While restoring text that had been stripped from committed pages, it saw short diff excerpts from other models' pages (`claude/`, `claude-opus-5/`, `codex/`, `luna-max/`), and diff excerpts of the skill file itself while checking its hash. It did not open any other model's page as a template. The skill was read in full only after the five no-skills pages were written.
- Standing platform and personal instructions applied throughout, as for every run.
- The checks below were written by the model being measured.

## Constraints followed

Standalone HTML with inline CSS and JavaScript, fictional local data, no backend and no build step. Google Fonts are the only remote resource; all illustrations are inline SVG. Every booking, message, post and payment is a local demonstration and says so on screen.

## Checks

No Playwright and no committed screenshots.

- Node parsed every inline script (`new Function`).
- Each page was loaded from a local server in the Claude desktop app's built-in browser with no console errors.
- A scripted pass at a 390 px viewport pressed every visible button and control, and confirmed there were no uncaught errors and no horizontal overflow.
- Targeted journeys: the FieldLedger tractor clash blocks saving and keeps the saved plan, then clears when the job moves to the free tractor; Quiet Practice rejects an amplified session in an acoustic-only room, a session running past a room's curfew, an end time before the start, and a booked hour; a valid booking confirms locally.
- Gallery previews in `assets/previews/*-opus55*.jpg` were captured with headless Microsoft Edge and inspected; they are gallery assets, not test evidence.

## Repairs after checks

Recorded separately from first-attempt output. The first attempts were not preserved as files.

1. Both FieldLedger pages started the day with two tractor clashes instead of the one intended; the spray job moved to the second tractor.
2. Both Second Sunday pages widened the mobile layout to 663 px because hidden radio inputs were positioned outside their labels.
3. The skill Quiet Practice chart clipped its last time label.
4. The no-skills FieldLedger report form gained a guard against submitting with no job selected; this was found by the scripted pass clicking inside a closed dialog, which users cannot reach.

Self-review fixes made while writing, before any check ran: a FieldLedger seed that didn't create the intended clash, focus handling in Second Sunday's message form, Quiet Practice showing a clash for hours just booked, a Second Sunday confirmation that lowercased a name, a packing list in the skill Borrowed Weekend with three items ticked for no reason, and field facts that lost their labels on mobile in the skill FieldLedger.