📰 Daily Digest — 2026-03-18

7 items | Business, DevTools, Open Source, AI, Misc

---

📋 Quick Summary

5 of my favorite new retention techniques that you may not have tried

Source: Elena’s Growth Scoop (Substack) · Category: Business · Link: Original

• Elena Verna shares five retention plays tested at Lovable, including cancellation-flow lite plans, involuntary churn messaging, daily free credits, rollover policies, and top-ups.
• Reported outcomes include a ~$5 hidden lite plan with ~10% take rate, a 50% relative payment-recovery lift (from ~20% to >30%), and a 7% retention improvement from top-ups.
• She emphasizes two retention laws: tests need multi-month observation, and the best retention decisions can look counterintuitive in short-term revenue views.

Will AI Kill Spreadsheets?

Source: a16z speedrun (Substack) · Category: Business · Link: Original

• The piece centers on Andrew Chen’s argument that many spreadsheet workflows are software in disguise and will increasingly move to code as AI lowers implementation barriers.
• It highlights a large debate thread with pro-code and pro-grid views, including claims that spreadsheets remain valuable for inspectability, intuition, and collaborative finance modeling.
• The emerging synthesis is convergence: AI-native tools may preserve the grid as an interface while shifting logic, testing, and scale concerns to software-style backends.

Lessons from Building Claude Code: How We Use Skills

Source: X (Thariq) · Category: DevTools · Link: Original

• Thariq describes internal Anthropic usage where hundreds of Claude Code skills are active, framing skills as full folders (scripts/assets/data/hooks), not just markdown snippets.
• The post maps recurring skill categories (knowledge, verification, data, workflows, scaffolding, quality, deployment, debugging, maintenance) with concrete examples.
• Key authoring advice includes encoding gotchas, using progressive disclosure, storing stable state via ${CLAUDE_PLUGIN_DATA}, and curating distribution through repos or marketplaces.

The Harness Is Everything: What Cursor, Claude Code, and Perplexity Actually Built

Source: X (Rohit) · Category: DevTools · Link: Original

• Rohit argues model performance is dominated by harness design (tools, context strategy, memory, and guardrails), citing SWE-agent’s 64% relative gain from interface improvements.
• The article details high-leverage ACI choices: capped search outputs (50), stateful file viewing (~100 lines with line numbers), and edit flows with immediate lint feedback.
• It connects these patterns to Anthropic and OpenAI operations, including structured docs, worktree isolation, explicit progress artifacts, and high-throughput PR pipelines.

get-shit-done: Context engineering and spec-driven system for coding agents

Source: GitHub · Category: Open Source · Link: Original

• get-shit-done positions itself as a lightweight context-engineering/spec-driven layer designed to reduce context rot in agentic coding workflows.
• The project supports multiple runtimes (Claude Code, Codex, Gemini CLI, Copilot, OpenCode, Antigravity) with installer flows for global/local and interactive/non-interactive setups.
• For Codex specifically, it uses skills-based installation under .codex/skills, and promotes command-driven workflows over heavyweight process overhead.

Why Garry Tan’s Claude Code setup has gotten so much love, and hate

Source: TechCrunch · Category: AI · Link: Original

• TechCrunch profiles YC CEO Garry Tan’s open-sourced Claude Code setup (“gstack”), shared after SXSW comments about intensive multi-agent usage.
• The article reports rapid adoption signals (nearly 20,000 GitHub stars and ~2,200 forks) alongside criticism that the package is mostly prompt workflow repackaging.
• Expert-model takes quoted in the story frame gstack as a sophisticated but non-magical orchestration pattern that can improve output correctness via role-structured flows.

Dupple tracking URL (destination unresolved)

Source: elink640.dupple.com · Category: Misc · Link: Original

• The inbox URL appears to be a tracking link and did not resolve to a readable destination in this environment.
• ⚠️ Fetch failed (Cloudflare anti-bot challenge/403 on available retrieval paths).
• Detailed analysis is deferred until a canonical destination URL is accessible.

---

📝 Detailed Notes

1. 5 of my favorite new retention techniques that you may not have tried

1. The article frames retention as the clearest quality signal for product-market fit.
   • Verna positions churn reduction as an always-on feedback loop, not a one-time campaign.
   • She says opportunities remain even for teams already advanced in retention analytics.
   • The practical angle is “recently tested techniques” with measurable lift, not theory.
2. Cancellation flow design starts with churn-reason intelligence and targeted save offers.
   • She reiterates that “too expensive” and “don’t need it anymore” are frequent churn drivers.
   • Lovable tested a hidden Lite plan in downgrade flow instead of only full-price-or-exit.
   • Pricing trials at $15/$10/$5 reportedly converged on $5 with ~10% take and better retention.
3. Involuntary churn recovery improved through in-product payment-failure prompts.
   • Email retries alone were described as insufficient for many users with failed payments.
   • Dashboard-level alerts reportedly drove a ~50% relative recovery increase.
   • The cited movement was from roughly ~20% update rate to above 30% in retry cohorts.
4. Credit mechanics are used as behavioral hooks and fairness levers.
   • The model gives 5 daily credits to free users and to paid users after credit exhaustion.
   • A rollover policy keeps purchased credits available while subscription remains active.
   • On full churn, accumulated credits are frozen (not erased) until resubscription.
5. The final takeaway is that retention optimization often looks wrong in the short run.
   • A top-up option reportedly added ~7% overall retention on a large paid base.
   • Verna argues multi-month measurement is required because early revenue can dip first.
   • She concludes that customer-experience generosity can compound long-term monetization.

---

2. Will AI Kill Spreadsheets?

1. The post anchors on Andrew Chen’s provocation that spreadsheet logic should move to code.
   • He argues grids often host “programs” lacking version control, testing, and modularity.
   • AI code generation is presented as collapsing the skill barrier for non-engineers.
   • The framing predicts a structural shift rather than a niche tooling trend.
2. The bullish migration case focuses on scale and software leverage.
   • The article references roughly 1 billion spreadsheet users globally.
   • It suggests even a 10% workflow shift could create a major wave of internal micro-apps.
   • Labs/startups are cited as already shipping AI-native spreadsheet-adjacent products.
3. Counterarguments center on cognition, inspectability, and finance-specific workflows.
   • Critics claim the grid remains unmatched for human interaction with tabular assumptions.
   • They highlight cell-by-cell auditing and rapid sensitivity exploration as core advantages.
   • PDF persistence is used as an analogy for why legacy formats can remain durable.
4. A useful middle model splits spreadsheet usage into two distinct jobs.
   • “Mini-software” trackers/dashboards are seen as the most replaceable category.
   • “Thinking tools” for model construction may persist longer due to cognitive fit.
   • This decomposition explains why participants can disagree while both being partly right.
5. The practical end-state discussed is convergence, not binary replacement.
   • Hybrid systems can keep grid UX while compiling logic into auditable software layers.
   • AI then acts as translator between intent, formulas, and production-grade execution.
   • The competitive frontier becomes trust, testability, and workflow ergonomics.

---

3. Lessons from Building Claude Code: How We Use Skills

1. The post begins by redefining what a “skill” is in operational terms.
   • Skills are described as folders containing instructions plus executable resources.
   • This enables agents to discover scripts, references, templates, and configs on demand.
   • Anthropic reports internal use at “hundreds of active skills,” implying broad utility.
2. It introduces a functional taxonomy to reduce design ambiguity.
   • Categories include domain knowledge, verification, analytics, workflow automation, and more.
   • The thread provides concrete examples like checkout verifiers and on-call runners.
   • A key warning is that multi-purpose skills become confusing and trigger unreliably.
3. Authoring guidance prioritizes high-signal constraints over generic documentation.
   • “Gotchas” are treated as the most valuable section because they encode real failure modes.
   • Progressive disclosure is encouraged via files such as references/api.md and templates.
   • Instructions should constrain outcomes while leaving room for contextual adaptation.
4. Durable state and executable helpers expand skill effectiveness over time.
   • Setup can be persisted via config files and resolved interactively when absent.
   • Historical logs/JSON/SQLite can provide memory across repeated workflow runs.
   • The stable storage path ${CLAUDE_PLUGIN_DATA} is noted for upgrade-safe persistence.
5. Sharing and governance are framed as social-technical systems.
   • Teams can distribute via repo-local skills or a curated plugin marketplace.
   • Promotion can be usage-led, with sandbox trials before formal marketplace placement.
   • Instrumentation through hooks is recommended to monitor trigger quality and adoption.

---

4. The Harness Is Everything: What Cursor, Claude Code, and Perplexity Actually Built

1. The thesis is that harness engineering outweighs model-choice obsession.
   • Harness is defined as full environment design: tools, memory policy, context format, guardrails.
   • The claim rejects reducing success to model version, token limits, or prompt wording alone.
   • It positions applied AI progress as a systems-design problem.
2. SWE-agent is used as empirical grounding for interface-first performance gains.
   • The article cites a 64% relative benchmark improvement from ACI over raw shell use.
   • Search output caps (50 results) are highlighted as protection against context flooding.
   • File viewing around 100 lines with state and line numbers is presented as a tested optimum.
3. Tight feedback loops prevent cascading failure in long agent runs.
   • Structured edit operations with immediate lint checks localize syntax failures quickly.
   • Without this, agents can chase downstream symptoms and waste context budget.
   • Context compaction is framed as necessary to remove stale observations over time.
4. Anthropic-style coding flows emphasize explicit state and clean handoffs.
   • Initializer/coder separation plus feature JSON with boolean passes fields reduce ambiguity.
   • Session closure expectations include clean state, progress updates, and recoverable checkpoints.
   • Browser automation (for example Puppeteer MCP) is treated as essential for true E2E verification.
5. OpenAI and broader ecosystem patterns point to scalable agent operations.
   • Cited metrics include roughly 1M lines of agent-written code and ~1,500 PRs in five months.
   • Repo-native docs, worktree isolation, and low-friction merge policy appear repeatedly.
   • A seven-layer stack framing treats coding agents as commoditized execution endpoints.

---

5. get-shit-done: Context engineering and spec-driven system for coding agents

1. The repository markets itself as a practical anti–context-rot layer.
   • It emphasizes reliability through context engineering and spec-driven execution.
   • The pitch targets builders who want outcome consistency without process ceremony.
   • Positioning is explicitly “lightweight workflow, heavy internals.”
2. Installation paths are broad, with runtime-specific targets and flags.
   • Primary onboarding is npx get-shit-done-cc@latest with guided runtime selection.
   • It supports global and local install modes to fit personal or team usage.
   • Non-interactive options are documented for CI, Docker, and scripted environments.
3. Multi-runtime support is a core design choice, not an afterthought.
   • Listed runtimes include Claude Code, Codex, Gemini CLI, Copilot, OpenCode, Antigravity.
   • Codex integration is skills-first, installed under skills/ structures in .codex.
   • Command conventions are adapted per runtime (/gsd:* or $gsd-* patterns).
4. The workflow philosophy prioritizes delegation with verification.
   • The author argues complexity should live in the system, not in manual operator rituals.
   • Claimed internals include orchestration, structured prompting, and state management.
   • Users are encouraged to run map/spec/task flows rather than ad hoc prompting.
5. Adoption signals and documentation breadth support ecosystem traction.
   • The README surfaces badges, social proofs, and enterprise-name trust indicators.
   • Licensing is MIT, with multilingual documentation and a detailed user guide.
   • The project also documents contribution-friendly local development installation.

---

6. Why Garry Tan’s Claude Code setup has gotten so much love, and hate

1. The story opens with Tan’s high-intensity public narrative around AI agents.
   • He describes sleeping about four hours while running multiple active agent projects.
   • He references “10 workers” and roughly three simultaneous coding efforts.
   • The interview tone is partly humorous but intentionally signals unusually high velocity.
2. gstack is presented as a reusable skill stack rather than a single prompt.
   • Tan open-sourced the setup on March 12 under an MIT-style sharing posture.
   • The package started with six opinionated skills and reportedly expanded quickly.
   • Workflow examples include CEO ideation, engineer implementation, and reviewer checks.
3. Distribution velocity became part of the product narrative.
   • The launch thread reportedly went viral and reached Product Hunt audiences.
   • TechCrunch cites nearly 20,000 stars and around 2,200 forks at publication time.
   • Fork activity suggests users saw immediate value in customization and remixing.
4. The backlash focused on novelty claims and authority amplification.
   • Critics argued the package resembled prompt bundles many practitioners already use.
   • A tweet about finding a subtle XSS issue triggered strong skepticism and ridicule.
   • Some commenters framed attention as influenced by YC status more than technical uniqueness.
5. External model commentary converged on a moderate interpretation.
   • ChatGPT/Gemini/Claude were quoted as viewing gstack as structured and useful, not magical.
   • The shared benefit is role-based decomposition that improves consistency and correctness.
   • The broader takeaway is that packaging operational patterns can itself be leverage.