| name | caseops-prd-execution |
| description | Use this skill for any CaseOps feature planning, implementation, product-gap analysis, module review, UX redesign, or PRD update. Forces every task to map to the unified PRD, current repo truth, source-data rules, user stories, and test IDs before work begins. |
CaseOps PRD Execution
This skill is mandatory for any CaseOps feature work, product planning, UX
revision, roadmap review, or module-gap analysis.
Read first
Before doing substantial work, read:
docs/PRD_CLAUDE_CODE_2026-04-23.md
docs/WORK_TO_BE_DONE.md
docs/PRD_COVERAGE_MOD_TS_2026-04-20.md
docs/STRICT_ENTERPRISE_GAP_TASKLIST.md
docs/STRICT_BUG_TASKLIST_2026-04-22.md when the task touches a bug,
regression, or reopen
Mandatory workflow
- Identify the journey ID or IDs affected.
- Identify the module ID or IDs affected.
- Inspect the current routes, services, pages, schemas, and tests before
claiming a gap or proposing a rewrite.
- Classify the work honestly against the current repo state.
- Map the task to user stories and test IDs from the PRD.
- Only then plan or implement the change.
- Update the PRD or the strict ledgers when the repo truth changes.
No-random-work rule
- Do not invent new modules, workflows, routes, or marketing claims that are
not grounded in the PRD.
- If the user asks for a genuinely new feature, update the PRD first or in the
same task.
- Do not mark a module "missing" just because one expected UX surface is absent
if the backend or an adjacent UI already exists.
- Do not mark a module "done" just because some code exists if the key journey
is still broken, partial, or unverified.
Vector and AI quality rules
For any retrieval, corpus, statute, tribunal, judge-intelligence, or
document-intelligence work:
- Production vector data must use the PRD's Voyage-based production path.
- Current production truth for CaseOps is Voyage
voyage-4-large on GCP,
not the historical BAAI/bge-small-en-v1.5 baseline.
- Corpus enrichment that materially affects retrieval quality must use the
high-reliability Opus-assisted normalization path or an explicitly approved
equivalent.
- Anthropic-backed enrichment, cleanup, or evaluation paths that are part of
the production corpus workflow must not be described as optional if the task
is about production readiness or production truth.
- Production retrieval must keep reranking enabled where the PRD requires it.
- No corpus slice may be called production-ready without the PRD's 4.8+/5
quality gate and benchmark evidence.
- Treat
bge-small, fastembed, or other local/offline models as dev,
smoke-test, or fallback paths unless explicit benchmark evidence says they
are the chosen production path.
- If older docs, notes, or spreadsheets still describe
bge-small as current
production, call that out as stale-doc drift instead of repeating it.
Bench-aware drafting rules
For any judge-profile, bench-match, appeal-drafting, recommendation, or hearing
strategy work:
- Inspect the current judge/court routes,
services/bench_matcher.py, drafting
templates, drafting prompts, and tests before claiming the feature is missing
or complete.
- Treat judge profile plus bench-match as foundation only. The feature is not
end-to-end unless drafting or recommendation generation consumes cited judge
or bench history.
- Add or update a task brief when bench-aware drafting scope changes. Current
queued brief:
docs/BENCH_AWARE_APPEAL_DRAFTING_TASKLIST_2026-04-24.md.
- No judge favorability scoring, win/loss prediction, reputation claims, or
uncited "judge tendency" language is allowed.
- Bench-aware output must use evidence phrasing: "in the indexed decisions
provided, the bench emphasized..." and must cite the supporting judgments.
- Weak or absent judge-history coverage must degrade to normal drafting with a
visible limitation note, not to invented strategy.
- Required tests include tenant isolation, weak-evidence fallback, no
favorability labels, citation verification, prompt-injection resistance, and
UI rendering of context quality.
UX rules
- Prefer simple, lawyer-friendly flows over feature density.
- One screen should have one obvious job and one obvious primary action.
- Error copy must be actionable and plain-language.
- Desktop-only proof is insufficient for surfaces lawyers will use on laptops,
tablets, or phones in real practice.
Documentation rules
- Keep
docs/PRD_CLAUDE_CODE_2026-04-23.md current when scope, sequencing, or
module status changes materially.
- Keep
docs/STRICT_ENTERPRISE_GAP_TASKLIST.md current when a feature task
exposes a broader platform or hardening gap.
- Keep
docs/WORK_TO_BE_DONE.md and docs/PRD_COVERAGE_MOD_TS_2026-04-20.md
aligned with live code when they drift.
Minimum output standard
Any substantial CaseOps feature answer should be able to state:
- affected journey IDs
- affected module IDs
- current status in the repo
- required user stories
- required functional tests
- required non-functional tests
- required security tests
If you cannot state those, you have not done enough context work yet.