repetitive-tasks

name	repetitive-tasks
description	Eliminate recurring work on gowkhtmltopdf — per-fixture screenshot/fix/regenerate loops, lint waves, benchmark re-runs, golden regeneration, temp diagnostic tests, re-taught conventions, and phase-checklist reconciliation. Use when a task looks like something already done on this project; update the skill or checklist instead of repeating the loop.

Repetitive Tasks

Evidence from the session-history scan (170 sessions, Aug 3–15):

• The identical per-fixture loop (screenshot PDF → list issues → fix → regenerate → commit) ran 5× verbatim in one session and ~10× in another, and the same goal was restarted in a brand-new session minutes later (Aug 12 fixes uncommitted → Aug 13 re-diagnosed from scratch).
• The same lint problems were fought in 4+ waves; make lint ran ~35× across two sessions; the layout lint-fix was done twice from scratch (batch-1 killed for git commands, batch-2 rebuilt the same renamer tool).
• paintCount (P2-12) was worked 3 separate times; P4-07 was attempted, reverted, then redone; same PDF-writer bug classes (zlib/FlateDecode, /Outlines ordering) were re-litigated across 4+ sessions.
• The 500-page benchmark was re-run 50+ times with results moving within a noise band; golden PDFs were regenerated in 5+ separate sessions; the same compliance PDFs were generated and committed twice.
• Temp diag_*_test.go files were created/run/deleted constantly (57 references in one session; the same file rm'd 9×).
• The em-dash rule, "commit go only / ignore PDFs", and branch-naming conventions were re-taught in 6+ sessions each.
• Phase checklists drifted from implementation, requiring explicit post-hoc "reconcile" passes.

Rules

1. Before repeating any loop, break the loop. If you have done the same diagnose→fix→regenerate cycle more than twice, stop and build the durable artifact: a saved QA/screenshot harness (see skills/avoidable-work/SKILL.md), a golden/screenshot-diff regression gate, or a make target. Loops are a signal the tooling is missing, not that effort is needed.
2. PDF fidelity loop. One canned flow: render fixture → read screenshot directly (multimodal) → one engine fix per root cause → regenerate → run the golden gate → commit. Do not re-list known issues; track open fixture issues in the canonical checklist instead.
3. Lint once per PR, never in waves. Run make lint before opening any PR and fix findings in the same change. Never run golangci-lint --fix or scripted renames as bulk passes (see skills/mistakes/SKILL.md). Config fixes (e.g. gofumpt/Go-version incompatibility → exclude in .golangci.yml) are done once, never re-diagnosed.
4. Golden artifacts have one source of truth. Golden PDFs are regenerated only by make samples from the committed generator; make test must never depend on files make samples rewrites. If both touch the same HTML, the conflict is a bug to fix, not a ritual to repeat.
5. Benchmarks are noise-aware. Fix the commit and dataset, define the acceptance gate once (e.g. 500-page < 2.5s), and stop when the gate is met. Do not chase movement within the noise band; do not re-run the full matrix for micro-optimizations without a measured bottleneck.
6. Diagnose with one durable harness. Keep a single debug harness (e.g. a fixture-driven test with -run filters) instead of creating/deleting diag_*_test.go per fixture. A compile error 22× on the same probe means the probe should live in the repo.
7. Conventions live in skills/AGENTS.md. Style rules (no em dashes), commit rules (go-only vs generated output), branch naming, and "no git commands" are read from AGENTS.md, not re-instructed. If a convention was re-explained to you, add it to AGENTS.md.
8. Checklists are live ledgers. Update phase-checklist rows ([x]) in the same change that implements them (see skills/phase-wise-checklist/SKILLS.md). Never leave reconcile as a follow-up pass.
9. Version sweeps are a single checklist. VERSION, internal/cli/help.go, CHANGELOG, docs, and frontend version strings change together in one commit, verified by TestCLIVersionMatchesVERSIONFile before committing.

Completion Handoff

• Any loop repeated 2+ times in this session has been converted into a script, harness, or checklist row.
• No temp diag_* files left in the tree.
• make lint clean before PR; no lint wave deferred to a follow-up.
• Phase-checklist rows match shipped work at session end.
• Benchmark/verification results recorded with commit + dataset so re-runs are comparable.