メニュー
Cross-session memory management. Accumulates patterns, pitfalls, and preferences. Review, search, prune, and export learnings so gstack gets smarter on your codebase over time.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Builds production-grade Claude Projects and CustomGPTs, improves AI output quality using meta-thinking prompts, and architects self-improving AI systems. Use proactively when a founder wants to build a reliable AI workflow, automate a repeatable task with AI, improve the quality of AI outputs they're getting, feels stuck with generic AI responses, or wants to build internal tools powered by Claude. Trigger for "build a Claude project", "my AI outputs are generic", "how do I make Claude more specific", "build an AI workflow", "automate this with AI", or "my prompts aren't working".
One-command review pipeline. Chains CEO review → design review → eng review automatically. Surfaces only taste decisions for PD approval. Everything else runs autonomously.
Performance baseline measurement. Captures page load times, Core Web Vitals, and resource sizes. Compare before/after on every PR.
Playbook for bootstrapped and early-stage SaaS founders based on proven case studies ($0 to $5M ARR with 6 people, and creator-to-business flywheel models). Use proactively when a founder is bootstrapped or capital-efficient, needs to grow without paid acquisition, is facing a churn crisis, wants to build a freemium or usage-based model, or asks "how do I grow without VC money". Also trigger for "PLG", "inbound-led outbound", "creator-led growth", "build in public", or "how did [bootstrapped company] grow".
Real browser automation for testing, QA, and demo verification. Provides Playwright-based browser control for navigating the live app, clicking elements, filling forms, taking screenshots, and verifying visual state.
Post-deploy monitoring. After release-engineer pushes to production, canary watches the live site for errors, performance regressions, and page failures.
| name | learn |
| description | Cross-session memory management. Accumulates patterns, pitfalls, and preferences. Review, search, prune, and export learnings so gstack gets smarter on your codebase over time. |
| used_by | ["ceo-thinking-partner","all agents"] |
Every session teaches something. This skill captures it so it compounds.
Primary store: docs/memory-summaries/learnings.md
After every significant event, append a learning:
| Trigger | Learning Type | Example |
|---|---|---|
| Bug found + fixed | PITFALL | "Database doesn't support JOIN — use subqueries" |
| PD design choice | PREFERENCE | "PD prefers card layouts over tables" |
| Agent blocker resolved | PATTERN | "Always run [BUILD_CHECK_COMMAND] before spawning qa-engineer" |
| Tier 3 gate hit | PROCESS | "Schema changes need migration file BEFORE backend code" |
| Test failure root cause | TECHNICAL | "Test runner needs explicit cleanup between tests with shared state" |
| Performance fix | PERFORMANCE | "[DOMAIN_ENTITY] list page needs pagination at >20 items" |
[LEARNING-NNN] [TYPE] — [date]
Context: [what happened]
Learning: [the lesson — one sentence]
Evidence: [file or commit reference]
Applies to: [which agents / which situations]
Review all learnings:
PD says "What have we learned?" → Read and summarise docs/memory-summaries/learnings.md
Search learnings: PD says "What do we know about [topic]?" → grep learnings.md for topic, present matches
Prune learnings: PD says "Clean up learnings" → Review each. Remove duplicates. Remove learnings that are no longer relevant (e.g., about a deleted feature). Present pruned list for PD approval before saving.
Export for new project: PD says "Export learnings for [project]" → Filter to PATTERN and PROCESS types only (skip project-specific PITFALLs). Write to a portable format.
At session START: CEO reads docs/memory-summaries/learnings.md (already in session start ritual)
At session END: CEO appends any new learnings from this session.
When an agent hits a problem: check learnings.md first. If a relevant learning exists, follow it instead of re-discovering.
EXPIRES: [date]