Skip to content

Latest commit

 

History

History
188 lines (127 loc) · 9.64 KB

File metadata and controls

188 lines (127 loc) · 9.64 KB

Pending Work — Asgard Skills

Tracking deferred items by priority. Updated 2026-04-07 after the rename + repo split.


🟡 Medium Priority — tw-ecommerce Skeleton Backlog

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).


🟡 Medium Priority — Skill Quality Improvements

1. QA Phase 2-3 P2/P3 Remediation

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-calendar
  • xborder-logistics
  • hum-socratic
  • meta-systems-thinking
  • stat-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 gotcha
  • soc-cognitive-bias — bias catalog too long; offload to references/
  • hum-socratic — over-teaches question types; missing "leading questions" anti-pattern
  • meta-systems-thinking — over-teaches feedback loop definitions
  • stat-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.

2. Strict eval of Tier 1 + 2 scripts ✅ DONE

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).

3. Tier 3 Scripts (heavier dependencies)

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.


🔴 High Priority — Phase 5: Plugin Bundling (Paused)

Plugin Bundle Strategy

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.

Available raw materials

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.

4 proposed plugin bundles (personas)

Plugin 1: asgard-ecommerce-analyst — Taiwan SMB e-commerce owner

  • 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

Plugin 2: asgard-taiwan-stock-analyst — Taiwan stock researcher ⭐ (start here, simplest)

  • 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

Plugin 3: asgard-taiwan-policy-researcher — Policy researcher

  • 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

Plugin 4: asgard-taiwan-finance-ops — Taiwan finance/accounting ops

  • 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

Repo layout decision: Option B → then C

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.

Recommended order

  1. Plugin 2 (stock) — simplest (1 MCP + 12 skills), validate the flow
  2. Plugin 1 (e-commerce) — most complex (4 MCPs + 15 skills), real-world case
  3. Plugin 3, 4 — follow

Why paused

The user wanted to first improve skills themselves (deterministic scripts, repo split, naming cleanup). Resume Phase 5 after this housekeeping.


⚪ Low Priority / Watch List

5. Subagent sandbox limitation ✅ RESOLVED

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.


Suggested next-step order (updated 2026-04-09)

Order Item Effort
1 Repo-wide broken-references sweep ✅ 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