# Three actual skill instruction runs — 2026-10-30

These original, non-sensitive fixtures were authored and executed by the Codex desktop agent in the implementation session for issue #8. The agent read the pinned upstream instructions and applied them directly using file, shell and browser tools. These are not automatic skill-loader benchmarks. The local Codex and Claude Code CLIs were available (0.161.0 and 2.1.283) but were not used to generate these results. Agent-authored workflow notes are not raw chat transcripts. The linked command outputs and screenshots are actual retained tool results.

Agent/model: Codex desktop, GPT-6 family; exact deployment identifier unavailable to the session. Run date: 2026-10-30 (Europe/Dublin). Environment: macOS/Darwin arm64, Node v26.8.1; browser checks used Playwright 1.64.1 and Chrome 156.0.8078.79. Existing account access was used; incremental cost was not measured. No private or production data was used.

## Pinned instructions and attribution

Upstream skill instructions remain the work of their respective authors. Read the immutable originals; this evidence bundle does not redistribute them:

- Emil Kowalski, [Design Engineering](https://github.com/emilkowalski/skills/blob/a81dbca07f2217b7103e510c0f13f8d5b017afc9/skills/emil-design-eng/SKILL.md).
- Jesse Vincent / Superpowers, [Systematic Debugging](https://github.com/obra/superpowers/blob/2ce8aa15d05bd57213e6f746f953aafb7ded7a8b/skills/systematic-debugging/SKILL.md). Also read [TDD](https://github.com/obra/superpowers/blob/2ce8aa15d05bd57213e6f746f953aafb7ded7a8b/skills/test-driven-development/SKILL.md), its [testing reference](https://github.com/obra/superpowers/blob/2ce8aa15d05bd57213e6f746f953aafb7ded7a8b/skills/test-driven-development/writing-good-tests.md), and [verification guidance](https://github.com/obra/superpowers/blob/2ce8aa15d05bd57213e6f746f953aafb7ded7a8b/skills/verification-before-completion/SKILL.md). Verification was also read from the installed local skill before the run; the pinned upstream version was checked afterward.
- Matt Pocock, [Grill Me](https://github.com/mattpocock/skills/blob/a496f5513d4a4cf889bc48e189e579f4b0bdf6aa/skills/productivity/grill-me/SKILL.md) delegates to [grilling](https://github.com/mattpocock/skills/blob/a496f5513d4a4cf889bc48e189e579f4b0bdf6aa/skills/productivity/grilling/SKILL.md). Both were read. The host lacked the Skill dispatch tool; directly reading and applying the dependency is an explicit adaptation, so this run is labeled partial.

Original fixture code, notes and screenshots in this directory are offered under the accompanying MIT license. No external media, fonts, brands or real member information are embedded. The fictional Fieldwork name conveys no affiliation. Upstream owners retain their own rights; their licenses govern their instructions. Claims here apply only to these bounded examples.

## Design reproduction

1. Read `design/request.md`, the pinned skill and `design/workflow.md` (motion choices and required Before/After table).
2. Open `design/before.html` and `design/result.html` in a modern browser. The latter is also editable source; `design/source.html.txt` is the same source as plain text. Toggle the summary and Save using pointer and keyboard. Refresh resets it.
3. To repeat automated checks, download `result.html` and `verify.cjs` into a folder named `design`. From its parent serve files with `python3 -m http.server 8768 --bind 127.0.0.1`. In `design`, run:

```sh
npm install --no-save playwright@1.64.1
node verify.cjs
```

The script uses installed Google Chrome (`channel: chrome`). It writes desktop/mobile screenshots and `verification.json` in the current directory; use a fresh copy to preserve the published originals. Set FIXTURE_URL if serving on a different port. Original generation used an isolated tools directory and NODE_PATH rather than adding dependencies to the application. Chrome DevTools connector initially failed because its profile was already running; Playwright launched a separate headless Chrome successfully. Five checks are enumerated in verification.json; they do not constitute an accessibility audit. The agent visually inspected desktop.png. No physical device or next-day review occurred.

## Debugging reproduction

Download `debugging/before.mjs`, `debugging/fixture.test.mjs` and `debugging/fix.patch` into a fresh folder. Read investigation.md for the four-phase reasoning. Run:

```sh
cp before.mjs fixture.mjs
node --test --test-reporter=spec fixture.test.mjs
# Expected exit 1: zero quantity costs zero fails (actual 1250, expected 0).
patch -p0 < fix.patch
node --test --test-reporter=spec fixture.test.mjs
# Expected exit 0: 4 tests pass.
```

The fixed file is retained as fixture.mjs for comparison. The suite is the entire fixture suite; it uses real functions, literal expectations and no mocks. red.txt and green.txt retain actual test output including timings. Publication removed only the temporary absolute directory in stack frames. This teaching bug was deliberately seeded by the same agent; it is not independent production evidence. No human edited the patch. Inputs outside the documented quantity contract were not tested.

## Requirements reproduction

Read `requirements/brief.md`. Invoke the pinned Grill Me/grilling instructions with an agent. Use `answers.json` as an explicitly scripted respondent, consuming the relevant round only after generating the questions. Each settled frontier unlocks the next round. Use the confirmation before generating the specification. The original agent followed three rounds, recorded in interview.md; the source Skill-tool call was replaced by direct reading. Environment facts were supplied in the brief, so no discovery delegation was needed.

Compare the resulting decisions with specification.md. Generative wording and frontier grouping may vary; exact output replay is not promised. Manual traceability links each product rule to Q1–Q6. No stakeholder validation or executable acceptance tests took place. Architecture, date-boundary details and operational guarantees remain explicit engineering refinement; this partial run does not prove a complete development-ready specification.

## Retention and publication

Raw generation lived in a separate temporary workspace; only selected artifacts were copied here. A pair of requirements notes briefly landed outside that temporary workspace during authoring and were moved into it before publication; no application-source changes were retained there. The website integration is separate from the fixture runs. No credentials, personal data or private absolute paths are published. SHA256SUMS records the selected evidence files (excluding itself) so local copies can be checked. Hashes detect accidental changes; they are not independent proof of agent behavior.
