Tracking deferred items by priority. Updated 2026-04-07 after the rename + repo split.
Added 2026-04-19 on branch feat/tw-ecommerce-domain. 26 skeletons with frontmatter + "when to use" only — bodies pending.
Design spec: docs/superpowers/specs/2026-04-19-tw-ecommerce-domain-design.md
Domain map: docs/domains/tw-ecommerce.md
Each follow-up branch fills in one layer:
| Branch | Skills to fill |
|---|---|
feat/tw-ecom-platform-skills |
tw-ecom-channel-strategy, tw-ecom-dtc-91app, tw-ecom-dtc-shopify-localization, tw-ecom-marketplace-shopee, tw-ecom-marketplace-momo |
feat/tw-ecom-payment-deep-skills |
tw-ecom-payment-tappay, tw-ecom-payment-ecpay, tw-ecom-payment-jkopay, tw-ecom-payment-dispute |
feat/tw-ecom-logistics-skills |
tw-ecom-logistics-cvs, tw-ecom-logistics-home, tw-ecom-logistics-cold-chain, tw-ecom-logistics-cross-border |
feat/tw-ecom-invoice-deep-skills |
tw-ecom-invoice-universalec, tw-ecom-invoice-carrier, tw-ecom-invoice-void |
feat/tw-ecom-compliance-skills |
tw-ecom-compliance-consumer, tw-ecom-compliance-product, tw-ecom-compliance-pdpa, tw-ecom-compliance-cross-border |
feat/tw-ecom-operations-skills |
tw-ecom-operations-promotion, tw-ecom-operations-pricing, tw-ecom-operations-line-oa, tw-ecom-operations-customer-service |
feat/tw-ecom-analytics-skills |
tw-ecom-analytics-ga4, tw-ecom-analytics-benchmarks |
Content-source discipline (per design spec §5.2):
- Platform / payment / invoice skeletons should wait for official docs + MCP snapshots before writing; avoid fabricating fee rates.
- Compliance skeletons must cite the law article (e.g.,
消保法 §19) + date. - Analytics benchmarks must cite industry source + year.
- Mark unsourced numbers as
TODO: verify with <source>— don't ship.
Deliberately excluded from the roadmap: tw-ecom-advertising, tw-ecom-reporting, tw-ecom-cyberbiz-integration (see design spec §6.3).
From the 28-skill quality audit (eval-workspace/qa-report.md in skill-template before split — recreate here if needed).
Weak Iron Laws (5 skills): some Iron Laws are truisms ("be specific", "look at the data first") instead of non-obvious constraints.
mkt-content-calendarxborder-logisticshum-socraticmeta-systems-thinkingstat-eda
Test for a good Iron Law: "Would a competent agent violate this rule WITHOUT being warned?" If no, the rule is too obvious and should be sharpened.
Over-teaching in 大學部 methodology skills: skills explain basic concepts (Nash equilibrium, IRAC steps, EDA workflow) the agent already knows. Compress definitions and expand gotchas/error-modes.
econ-game-theory— over-teaches dominant strategy / NE basics; missing mixed-strategy gotchasoc-cognitive-bias— bias catalog too long; offload to references/hum-socratic— over-teaches question types; missing "leading questions" anti-patternmeta-systems-thinking— over-teaches feedback loop definitionsstat-eda— over-teaches workflow; missing Simpson's paradox, data leakage gotchas
Populate examples/ directories (systemic): currently every examples/ directory exists only with .gitignore. CLAUDE.md says examples are "always required." Inline examples in SKILL.md partially compensate but the standalone files are still missing.
Completed 2026-04-09. Two methodologies run; see eval/eval-report.md
for full results and comparison.
Methodology B (skill ablation — primary) tally (20/20 × Sonnet 4.6):
| Verdict | Count | Skills |
|---|---|---|
| correctness ⭐ | 5 | algo-risk-altman-z, algo-price-elasticity, algo-rank-elo, algo-sc-eoq, algo-seo-tfidf |
| both ⭐⭐ | 1 | ecom-rfm-analysis |
| speed | 3 | biz-dcf, mkt-ab-testing, algo-sc-newsvendor |
| neither | 10 | rest |
| inconclusive | 1 | algo-ecom-bm25 (skill bug was fixed mid-run) |
Hit rate: 9/20 = 45%. The 10 "neither" skills are removal candidates but should be re-validated with 3–5 replication runs before any removal (N=1 sampling showed real verdict variance on at least 2 cases).
Side effects: eval surfaced and fixed skill quality issues in
algo-risk-altman-z (broken references + missing Phase 1.5 variant-selection gate),
algo-ecom-bm25 (undocumented stop-word removal), biz-financial-ratios (debt
definition ambiguity), and decimal-vs-percent conventions in 3 Output Format
sections (grad-capm, biz-dupont, biz-financial-ratios).
Scripts that require non-stdlib libraries:
- Needs numpy/networkx/statsmodels:
algo-seo-pagerank(sparse matrix iteration)algo-net-centrality(networkx for graph metrics)algo-net-community(Louvain via python-louvain)algo-forecast-arima(statsmodels)algo-forecast-prophet(Meta Prophet)algo-rank-trueskill(Bayesian message-passing)
- Needs lexicon data:
algo-social-sentiment(VADER lexicon)
Decision gating: complete Tier 1+2 strict eval first to confirm scripts add value on messy data; only then invest in Tier 3.
Plugin = Persona × (Data + Methodology). MCPs and Skills are raw ingredients. A plugin is a curated bundle of MCPs + ~10-20 skills targeting ONE specific user workflow. Value comes from CURATION, not completeness.
MCPs (8 under asgard-ai-platform, excluding mcp-template and asgard-mcp-server):
| MCP | Tools | Domain |
|---|---|---|
mcp-tdcc |
8 | Taiwan TDCC securities custody open data |
mcp-tw-ly |
— | Taiwan Legislative Yuan open data |
mcp-twfood |
6 | Taiwan agricultural wholesale market data |
mcp-shopline |
19 | Shopline e-commerce platform |
mcp-91app |
17 | 91app e-commerce platform |
mcp-ezpay-einvoice |
7 | Taiwan ezPay e-invoice |
mcp-universalec-e-invoice |
27 | Universal EC e-invoice |
mcp-newebpay |
8 | NewebPay payment gateway |
(mcp-cyberbiz is an empty repo — skip.)
Skills: 263 in this repo across 21 categories.
- MCPs:
mcp-shopline+mcp-91app+mcp-newebpay+mcp-ezpay-einvoice - Skills (~15):
ecom-rfm-analysis,data-cohort-analysis,ecom-analytics,ecom-inventory-health,biz-unit-economics,biz-cac-ltv,biz-pricing-strategy,ecom-promo-roi,algo-rec-cf,algo-price-elasticity,algo-forecast-arima,tw-einvoice-guide,tw-payment-integration,tw-tax-basics
- MCPs:
mcp-tdcc - Skills (~12):
tw-stock-analysis,fin-earnings-summary,biz-dcf,biz-financial-ratios,biz-dupont,data-financial-analysis,fin-modeling,grad-capm,grad-fama-french,grad-emh,grad-behavioral-finance
- MCPs:
mcp-tw-ly+mcp-twfood - Skills (~15):
grad-policy-streams,grad-governance,soc-policy-analysis,grad-case-study,grad-grounded-theory,grad-narrative,grad-systematic-review,grad-mixed-methods,hum-source-criticism,hum-discourse,hum-historical-analogy,hum-critical-thinking,meta-systems-thinking,meta-scenario-planning
- MCPs:
mcp-ezpay-einvoice+mcp-universalec-e-invoice+mcp-newebpay - Skills (~10):
tw-einvoice-guide,tw-tax-basics,tw-fintech-compliance,tw-startup-legal,data-financial-analysis,biz-financial-ratios,biz-breakeven,ops-contract-review,law-contract,ops-meeting-minutes
Phase 5a (prototype): Add bundles/ directory in skill-template (NOT this repo). Each bundle is a YAML manifest listing its MCPs + curated skills (referencing this repo by path or git submodule). Low risk, fast validation of curation logic.
Phase 5b (production): Write a build script that generates independent plugin repos from bundle YAML files. Publish as separate asgard-<plugin-name> repos.
- Plugin 2 (stock) — simplest (1 MCP + 12 skills), validate the flow
- Plugin 1 (e-commerce) — most complex (4 MCPs + 15 skills), real-world case
- Plugin 3, 4 — follow
The user wanted to first improve skills themselves (deterministic scripts, repo split, naming cleanup). Resume Phase 5 after this housekeeping.
Resolved 2026-04-08 by adding Bash(python *) and Bash(python3 *) to
.claude/settings.json permissions. Subagents can now execute Tier 1+2 scripts
without per-call approval. Reproducible across the project for anyone who clones it.
| Order | Item | Effort |
|---|---|---|
| ✅ Done 2026-04-09 (314/314 files generated) | ||
| 2 | QA P2-P3 fixes (weak Iron Laws, over-teaching, empty examples/) |
Medium |
| 5 | Eval replication: 3–5 independent runs per case to stabilize the 10 "neither" verdicts before removal | Medium |
| 6 | Phase 5b: build script + publish independent plugin repos | Large |
| 7 | Tier 3 scripts (only if plugin scenarios need them) | Large |