Frontend Design Improvements Loop

Frontend Design Improvements Loop

Use this skill to run end-to-end benchmark iterations that improve frontend-design behavior through instruction tuning, not model weight tuning.

Scope

• Repository: improved-frontend-skills-for-gpt
• Canonical prompt: repo root prompt.md (same prompt for every version)
• Active version workspace: experiments/version-X/
• Preferred raw baseline skill seed: global $frontend-design at /home/waishnav/.agents/skills/frontend-design/SKILL.md
• Primary references:
  • research/targetted-designs/
  • research/theo-screenshots-2k-clean/opus45_with_skill/
  • research/theo-screenshots-2k-clean/opus_iterations/

Read references/opus_targets.md, references/mutation_axes.md, and references/experiment_topologies.md before drafting a new version.

Non-Negotiables

1. Never modify prompt.md unless explicitly asked.
2. Always run Codex in the target version folder (-C experiments/version-X/...).
3. Keep one mutation hypothesis per version.
4. Keep each version self-contained (SKILL.md, t4-canvas/, README.md, CRITQUES.md, screenshots/).
5. Do not delete previous versions.
6. New versions must be isolated by default: do not inherit prior t4-canvas implementation unless explicitly requested.
7. Run one active experiment at a time (strict sequential loop; no parallel sibling branches).
8. Carry forward only one prioritized critique into the next mutation.
9. Keep SKILL.md portable: no experiment/version labels or repo-specific benchmarking terms.

Workflow

1) Choose baseline and one mutation hypothesis

• Baseline is the latest completed experiments/version-*.
• Pick one mutation axis only (see references/mutation_axes.md).
• Read the previous version CRITQUES.md and pick exactly one highest-impact issue to target.
• Write the hypothesis in the new version README before implementation.

2) Create the next version

• For true raw start (no version baseline), use:
  • scripts/new_fresh_version.sh <new-version-dir>
• Use scripts/new_version_from_previous.sh <previous-version-dir> <new-version-dir>.
• Default behavior is isolated: copy prior SKILL.md, create fresh empty t4-canvas/.
• Legacy mode only when explicitly requested: --copy-app.
• For raw baseline seeding from global frontend skill: use --skill-source global.

2b) Keep the loop sequential

• Do not create sibling versions in the same iteration.
• Do not run multiple mutation theses in parallel.
• Complete critique and keep/drop decision for the current version before creating the next version.

3) Tune frontend-design skill instructions

• Edit only experiments/version-X/.agents/skills/frontend-design/SKILL.md.
• Keep constraints auditable and measurable.
• Avoid vague language; use explicit guards and pass/fail criteria.
• Apply one mutation change set only for this version.
• Translate critique findings into generic rules; do not write benchmark- or version-specific wording inside SKILL.md.

4) Run Codex headlessly with canonical prompt

• Use scripts/run_headless_iteration.sh <version-dir> <repo-root>/prompt.md [run-label].
• Optional reliability args: [max-attempts] [retry-delay-sec].
• This forces codex exec to run with cwd inside version folder and logs artifacts.
• Prefer short, restartable runs over one giant run.

5) Capture screenshots for /1.. /5

• Start the version app in t4-canvas/.
• Use Playwright CLI to capture full-page screenshots for routes /1 to /5 in desktop/tablet/mobile profiles.
• Preferred command:
  • scripts/capture_screenshots.sh <version-dir> [base-url]
• Save under experiments/version-X/screenshots/<route>/{desktop,tablet,mobile}.png.

6) Critique and score

• Score against rubric in references/scoring_rubric.md.
• Compare against Opus reference sets (not generic web quality).
• Write critique notes to experiments/version-X/CRITQUES.md after every run.
• Record critique in a structured way:
  • issue
  • evidence
  • likely root-cause rule
  • proposed generic fix
• Record:
  • wins
  • regressions
  • next mutation

7) Decide next action

• If net gain: keep mutation and continue.
• If mixed: keep only if rubric delta is positive on target dimensions.
• If regression: revert mutation in next version and try a different axis.
• In all cases, pick one next mutation only.

Reliability Pattern For Long Runs

• Design for 3-7 minute chunks per run.
• Do not require tmux by default; run direct codex exec with artifact checkpoints.
• Persist outputs every run:
  • artifacts/<run-label>/events.jsonl
  • artifacts/<run-label>/stderr.log
  • artifacts/<run-label>/final.md
• If interrupted, continue with codex exec resume --last and continue from latest checkpoint.

Deliverable Contract Per Version

Each version must include:

• experiments/version-X/.agents/skills/frontend-design/SKILL.md (mutation applied)
• experiments/version-X/README.md with:
  • hypothesis
  • exact mutation
  • rubric score delta
  • next step
• experiments/version-X/CRITQUES.md with expected-vs-output critique notes
• experiments/version-X/screenshots/1/{desktop,tablet,mobile}.png ... screenshots/5/{desktop,tablet,mobile}.png

Scripts

• scripts/new_fresh_version.sh
• scripts/new_version_from_previous.sh
• scripts/run_headless_iteration.sh
• scripts/capture_screenshots.sh