AI as the interface for knowledge work: external research (2025–2026)

All sources below were opened unless marked. [partial] means only a preview or abstract was readable, usually because of a paywall.

1. Conversation as interface; apps as storage

• Jamin Ball, "Long Live Systems of Record", Clouded Judgement, 12 Dec 2025. https://cloudedjudgement.substack.com/p/clouded-judgement-121225-long-live. Argues that agents separate the interface for work from the source of truth, and that this raises the bar for systems of record rather than killing them. Supports and sharpens our thesis: the storage layer matters more, not less.
• Steven Lin, "SaaS Is Becoming Long-Term Memory for AI Agents", 15 Jan 2026. https://stevenlin.substack.com/p/saas-is-becoming-long-term-memory. Proposes a hierarchy: human → agent → local files as short-term memory (CLAUDE.md) → SaaS as long-term memory. Choose apps for complete APIs, clean data models and audit trails. Strongly supports "apps are storage" and gives a useful two-tier model.
• Matt Webb, "Headless", 18 Apr 2026. https://interconnected.org/home/2026/04/18/headless. Simon Willison's commentary on it, 19 Apr 2026: https://simonwillison.net/b/9416. Webb argues services will go headless: CLIs and MCP for personal AIs, with the GUI kept mainly for brand. CLIs are composable and have a smaller security surface. Willison adds Salesforce "Headless 360" and the threat to per-seat pricing. Supports our point about choosing tools for how well AI can read and write them.
• Karpathy, "Software Is Changing (Again)", YC talk, June 2025. [Not opened directly; I read secondary summaries in search results.] Makes the "build for agents" / llms.txt point. Supports our thesis.

2. Practitioners using agents for non-coding work

• Lenny Rachitsky, "Everyone should be using Claude Code more", 14 Oct 2025. https://www.lennysnewsletter.com/p/everyone-should-be-using-claude-code [partial: items 1–15 of 50]. Treats Claude Code as a local agent. Examples: a weekly self-review command that compares journal entries with commits and flags gaps between intentions and actual work; voice notes clustered into themes; call synthesis checked against a hypothesis list. Supports our "plan through the AI" principle.
• Teresa Torres (via ChatPRD / How I AI), 19 Jan 2026. https://www.chatprd.ai/how-i-ai/teresa-torres-claude-code-obsdian-task-management. Uses a /today command that builds a dated to-do list (due today, overdue, in progress). Each task is a Markdown file with YAML metadata. Context is kept in small modular files with index files. At the end of a session she asks "what did you learn that should be documented?" She left Trello for Markdown. Directly matches our "what's on today" practice and "apps are storage".
• ballred/obsidian-claude-pkm (GitHub, 2026, v3.1). https://github.com/ballred/obsidian-claude-pkm. Provides /daily, /weekly, /monthly and a /review router. Includes a per-project CLAUDE.md and a session-start hook that shows the "ONE Big Thing" and days since the last review. A ready template for our weekly-review practice.
• Anthropic, Claude Cowork (launched Jan 2026). https://claude.com/blog/cowork-research-preview. Brings Claude Code to non-developers through folder access, scheduled tasks and plugins. Anthropic said users "quickly began using it for almost everything else" (quoted via search results). Supports our thesis: the vendor is betting on this.

3. Shared skills, prompts and conventions

• Anthropic, "Agent Skills", 16 Oct 2025 (updated Dec 2025). https://claude.com/blog/skills. Skills are folders of instructions and scripts that load only when relevant. They work across the Claude apps, Claude Code and the API, and can be shared via git. This is the delivery mechanism for our processes, so a guide practice could ship as a skill.
• AGENTS.md. https://agents.md/. A cross-vendor "README for agents" (OpenAI Codex, Cursor, Google Jules and others), now with the Linux Foundation's Agentic AI Foundation. Extends our approach: write CLAUDE.md conventions so other tools can read them too.
• rb17080/session-handoff. https://github.com/rb17080/session-handoff. A HANDOFF.md with these sections: did / shipped / failed or stuck / learned / where things stand / next session start here. It pulls git data before writing, is rewritten each session and is gitignored, and promotes durable facts to memory. Directly borrowable.

4. Progress logging and handoff

• Steve Yegge, Beads. Repo: https://github.com/steveyegge/beads. Introductory post (Oct 2025): https://steve-yegge.medium.com/introducing-beads-a-coding-agent-memory-system-637d7d92514a [returned 403; content seen only in search excerpts]. Argues that Markdown plans rot: he found 605 plan files "in varying stages of decay". Agents need structured, queryable tasks with dependencies, plus bd ready to show unblocked work. Now backed by Dolt, with bd prime and bd remember. Supports using Beads for tasks. It also challenges any reliance on prose plans for live state.
• paddo.dev, "Beads: memory for coding agents", 15 Dec 2025. https://paddo.dev/blog/beads-memory-for-coding-agents. Describes "landing the plane": at session end, update issues, sync, clean up git, remove debugging leftovers, and generate the next session's prompt from the top unblocked work. It also notes Beads is "not a planning tool" or Jira replacement. Extends our approach: this is the end-of-session routine we want.
• Justin Young (Anthropic), "Effective harnesses for long-running agents", 26 Nov 2025. https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents. The pattern: a progress log, a JSON feature list with a passes flag (JSON because models are less likely to tamper with it than with Markdown), and a fixed start routine (read the log and git, pick the top item, check that things still work). Work one item at a time, leave things clean, commit. Supports our approach, and gives a rationale: structured data for state, prose for narrative.
• Anthropic, "Effective context engineering for AI agents", 29 Sep 2025. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents. Covers structured note-taking (NOTES.md, to-do lists), compaction and sub-agents, matched to task type. Supports our approach and is the conceptual frame for "processes designed for humans and AI".

5. Voice-first working

• Chris Blattman, "Voice Dictation" (undated, ~early 2026). https://claudeblattman.com/essentials/wispr-flow/. He "basically never types anymore". The main benefit is starting, not speed. He dictates rough prompts into Claude Code and expects about a week of training. Supports "talk, not type". His framing (lowers the blank-page barrier) is better than the speed claims.
• Karpathy's "vibe coding" tweet, 2 Feb 2025, reproduced on https://superwhisper.com/vibe-coding (a vendor page). He "barely even touches the keyboard" by talking via SuperWhisper. Supports our principle, but the same tweet endorses "Accept All" without reading diffs, which is the opposite of our steer principle.

6. Counterpoints

• METR, 10 Jul 2025. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/. Experienced developers were 19% slower with AI yet believed they were 20% faster. The page now flags these results as superseded by a Feb 2026 follow-up, which I did not open. Challenges felt productivity: measure, don't assume.
• Niederhoffer et al., "AI-Generated 'Workslop' Is Destroying Productivity", HBR, 22 Sep 2025. https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity [partial; figures from secondary coverage]. About 40% of workers had received polished but hollow AI output in the past month, and recipients trusted senders less. Challenges us: AI-written plans and updates handed to colleagues carry a cost.
• Kosmyna et al., "Your Brain on ChatGPT", arXiv, Jun 2025 (revised Dec 2025). https://arxiv.org/abs/2506.08872. People who wrote with an LLM showed the lowest brain engagement, the least sense of ownership, and poor recall of their own essays. Small study, preliminary. Supports our "you still think" caveat.
• Ball / Salesforce TDX coverage: SaaS vendors push back on "SaaS is dead" (the Salesforce TDX coverage was seen only in search results). This is a nuance for us: the app does not vanish. It becomes the system of record that agents talk to.

Patterns to borrow

1. Landing the plane: a named end-of-session routine. Update Beads, record what failed, write "next session: start here", and generate the resume prompt from the top unblocked work.
2. /today and /weekly commands (Torres, obsidian-claude-pkm): build the day's view from stored tasks, and run a weekly review that compares intentions with what actually happened.
3. Two-tier memory (Lin): local context files are short-term memory and the apps are long-term memory. Keep context modular and indexed (Torres) rather than in one giant CLAUDE.md.
4. "What did we learn?": at session end, ask the AI to propose updates to durable docs, and promote durable facts out of the transient handoff.
5. Structured state, prose narrative (Anthropic harness, Yegge): live status goes in Beads or JSON fields. Markdown holds rationale and decisions.
6. Ship processes as Skills and write conventions so AGENTS.md-compatible tools can read them too.

What would sharpen our principles

• "Apps are storage" → "apps are systems of record." Choose them for API/CLI/MCP quality, data model clarity and audit trail, not UI. Name the criteria.
• "Plan through the AI" needs a guard against Markdown plan rot. Live state belongs in the tracker, never in prose plans. We already do this with Beads, so state the reason.
• "You still think and steer" should cite the evidence (METR's perception gap, the cognitive-debt study) and add a social rule against workslop: you own anything the AI drafts that you send to others, and you read it first.
• "Talk, not type": lead with "voice lowers the cost of starting and of giving rich context", not with speed claims, which come from vendors.
• Record failures: a handoff should log dead ends as well as progress. This is missing from most of our descriptions.

For: How we use AI guide; brief.