2026-10-11-01 · T2 · Reasons · Analytical · "Agents are a 10-hour research habit for every buyer. Who loses?"
Reader: data and product leaders whose growth runs on distribution, not product quality
Promise: which business models lose when every buyer has an agent doing research (upsells, coupons, forgotten subscriptions, paid placement), and why quality data becomes the moat
Source: Patrick Collison, X post 2108706907859116280 (2026-10-09), shared by the founder 2026-10-11. His side point: data is "more valuable simply as a complement to cheaper intelligence."
Experience: NEEDS-BEN optional (the angle from his vibe-manufacturing / small-maker bet)
Jev: not run (manual seed)
2026-10-11-02 · T2 · Mistakes · Analytical · "Agents will reward the best product. Only if they can read it."
Reader: product and data leaders at brands selling online
Promise: why catalog listings, APIs/MCP and structured product facts are the new shelf placement, even if agents favor quality
Source: pairs Collison (X 2108706907859116280) with gaut on Meta Muse (X 2108958723725049987, 2026-10-10: 25 prompts, seller trust beat price, catalog-first, gives up on unreachable sites; n=25 anecdotal; author sells F500 consulting). Both shared by the founder 2026-10-11.
Experience: none required
Jev: not run (manual seed)
2026-10-11-03 · T1 · Frameworks · Analytical · "Two engineers, neither writes code. The scarce skill is the review layer."
Reader: engineering leaders reorganizing teams around agents
Promise: the pirate/architect split (0→1 vs 1→∞), and why the architect's real job is the eval and review layer
Source: Chintan Turakhia (Coinbase), X 2108981288317468720 (2026-10-10), slide "Two archetypes. Neither writes code." Shared by the founder 2026-10-11. "100x" claim has no data, so do not repeat it.
Experience: NEEDS-BEN (his own pirate+designer role with the studio: 5 packet calls on 10-11; phData architect framing)
Jev: not run (manual seed)
2026-10-11-04 · T1 · Mistakes · Analytical · "Cut headcount the day AI saves time, and you've seen the last good AI idea from your team."
Reader: executives and data leaders rolling out AI inside their companies
Promise: why AI adoption runs on bottom-up ideas, and why converting the first efficiency gain into layoffs shuts that pipeline off (the Toyota kaizen rule: "never cut people because improvement cut the hours")
Source: @blackright0, X 2109153023855616261 (2026-10-11, Japanese, 489k views), shared by the founder 2026-10-11. The no-layoff kaizen rule is well documented in lean literature (Ohno; NUMMI employment-security pledge); cite those, not the post.
Experience: NEEDS-BEN (what phData clients do with freed hours; where he has seen adoption stall)
Jev: not run (manual seed)
2026-10-28 founder ideas (iMessage 23:47Z-00:36Z, his own; quotes verbatim)
2026-10-28-01 · AI chickens · "What can my two cents buy nowadays?"
Founder: "What did we used to buy, outsource, or skip because each unit cost too much? I might also add because we didn't have the in-house expertise for... intelligence is getting cheap and you can use it to sort through a lot of problems you wouldn't have been able to tackle before."
Contrast: "what happens every week" use cases = incremental evolution; chickens = new processes and products.
Analogy: a vertically integrated quick-serve chain adds private-label eggs that also feed its own bakery (generic; client person removed per founder: "We can certainly take Tom out of the story").
Story: [[DESIGN-scribble-works|Scribble Works]]. His wife made games as "a labor of love"; now "a game studio that allowed us to make custom games on demand for pennies."
Status: founder 2026-10-28 00:40Z: "I think we tease that out into multiple posts. Too many themes in one thread." Split:
01a: the question itself. "Every week" use cases = faster today; buy/outsource/skip/no-expertise = new work. Own the consultant irony.
01b: the chicken. Eggs on the shelf AND in the bakery: one capability, two jobs (data layer as internal input + product).
01c: Scribble Works. "What can my two cents buy nowadays? Apparently some fun kids games!"
2026-10-28-02 · Use cases vs jobs to be done
Founder: "what are 'use cases'? How do you know what a good use case is? Is 'job to be done' better than use cases?"
Ray framing (Ray-selected, awaiting his read): a use case is the vendor's view, a job to be done is the buyer's; good-candidate test = weekly, painful, one owner, verifiable output.
2026-10-28-03 · Crawl, walk, run: get yourself slowly out of the loop
Founder: "Don't try to jump straight to durable agents proactively processing end to end workflows. AI-assisted is a better way to start and then move to AI-first. Get yourself slowly out of the loop."
Ties to post 1's "AI-native" (series: assisted → first → native).
2026-11-13 Claude fundamentals tip bank (founder direction: generic fundamentals, consistency first)
Series hashtag candidates (his pick): #ClaudeTips (broadest, most likely already in use), #ClaudeFundamentals (less likely crowded, matches the brief), #ClaudeBasics. Crowding could not be measured from outside LinkedIn; check the hashtag feed before picking.
2026-11-13-01 · T1 · Tip · Practical · "Did you know Claude follows a rule better when you say why?"
Tip: A bare rule like "NEVER use ellipses" works less well than the rule plus its reason. Anthropic's own example: say the output will be read aloud by a text-to-speech engine that can't pronounce ellipses. Claude generalizes from the reason, not just the rule.
Gotcha: In Anthropic's example, the all-caps "NEVER" on its own is the less effective version; shouting is not a substitute for the reason.
Snippet: "Your response will be read aloud by a text-to-speech engine, so never use ellipses."
Experience: NEEDS-BEN optional (a rule of his own that only stuck once he added the reason)
2026-11-13-02 · T1 · Tip · Practical · "Did you know 3-5 examples beat a paragraph of instructions?"
Tip: Show Claude what good looks like. Anthropic recommends 3-5 examples that mirror your real use case, wrapped in tags (several in ) so Claude can tell them apart from your instructions.
Gotcha: Make them diverse; near-identical examples teach Claude patterns you never meant (same length, same opening, same format).
2026-11-13-03 · T1 · Tip · Practical · "Did you know XML tags make long prompts easier for Claude to parse?"
Tip: When a prompt mixes instructions, context, examples and data, wrap each in its own tag: , , . Claude reads the boundaries and mixes them up less often.
Gotcha: There is no magic tag list; what matters is descriptive names used consistently across your prompts.
2026-11-13-04 · T1 · Tip · Practical · "Pasting a long document? Put your question at the end."
Tip: Put long documents near the top of the prompt and your question and instructions below them. Anthropic reports queries at the end can improve response quality by up to 30% in tests, especially with multiple documents.
Gotcha: With several documents, wrap each in tags with a so Claude can tell which one it is quoting.
2026-11-13-05 · T1 · Tip · Practical · "Did you know a Claude Project is also a search boundary for your past chats?"
Tip: A Project gives every chat in it the same uploaded knowledge and instructions. On paid plans Claude can also search your past chats, but inside a project it only searches that project's conversations, and project memory stays separate from other projects.
Gotcha: A decision made in a chat outside the project won't be found from inside it. Put anything load-bearing in the project knowledge rather than relying on search.
Experience: NEEDS-BEN optional (the project he keeps for one recurring kind of work)
2026-11-13-06 · T1 · Tip · Practical · "Did you know you can read and edit what Claude remembers about you?"
Tip: In the Claude app, memory is a list you can open, edit and delete under Settings. Each Project keeps its own memory, separate from other projects.
Gotcha: Incognito chats don't feed memory or chat search, so use one when you don't want a conversation to stick.
2026-11-13-07 · T1 · Tip · Practical · "Did you know a Claude Skill costs about 100 tokens until it's needed?"
Tip: A Skill is a folder with a SKILL.md. Until it triggers, Claude only holds its name and description (about 100 tokens). The full instructions load when a request matches the description, so the description has to say what it does and when to use it.
Gotcha: Custom Skills don't sync across surfaces; one uploaded to the Claude app isn't in the API or Claude Code until you add it there too.
Snippet: description: Extract tables from PDFs. Use when the user mentions PDFs or forms.
Experience: NEEDS-BEN optional (one skill he actually uses weekly)
2026-11-13-08 · T1 · Tip · Practical · "Did you know /init writes your first CLAUDE.md for you?"
Tip: In Claude Code, /init scans the repo and drafts a CLAUDE.md with build commands, test steps and conventions. Then trim it: the docs say aim for under 200 lines, because longer files cost context and get followed less.
Gotcha: @path imports organize a long CLAUDE.md but don't shrink it; imported files load in full alongside it.
2026-11-13-09 · T1 · Tip · Practical · "Did you know CLAUDE.md is a suggestion, not a guardrail?"
Tip: Claude treats CLAUDE.md as context, not enforced configuration. If something must always happen (format after every edit) or must never happen (a risky command), use a hook: a shell command Claude Code runs at a fixed point, every time.
Gotcha: A hook only fires for events its matcher covers (Edit|Write here); type /hooks to confirm it's registered under the right event.
Experience: NEEDS-BEN optional (a rule he moved from instructions to a hook)
2026-11-13-10 · T1 · Tip · Practical · "Press Shift+Tab before Claude Code touches a single file."
Tip: Plan mode lets Claude read the code and propose a plan, but it makes no edits until you approve. Press Shift+Tab until the status bar says plan mode, or start with the flag below.
Gotcha: The plan is only as good as your review of it; a wrong direction caught here is cheap, caught after 40 edits it isn't.
2026-11-13-11 · T1 · Tip · Practical · "Switching tasks in Claude Code? Type /clear first."
Tip: Claude Code resends the whole conversation with every request, so old context from an unrelated task costs tokens and muddies answers. /clear starts fresh; /compact keeps a summary, and you can say what to keep.
Gotcha: /compact itself reads the whole conversation, so it's a big request; /clear costs nothing. Use /rename first if you might come back.
Snippet: /compact Focus on the schema decisions and failing tests
2026-11-13-12 · T1 · Tip · Practical · "Did you know a subagent can read 50 files and hand back one summary?"
Tip: Subagents run in their own context window. Ask Claude Code to "use a subagent to investigate X" and the file reads and logs stay there; only the findings come back to your main session. Define reusable ones as Markdown files in .claude/agents/.
Gotcha: It keeps your main context clean, but the subagent's own requests still count toward your usage.
Snippet: use a subagent to investigate how our auth system handles token refresh
2026-11-13-13 · T1 · Tip · Practical · "Sending the same long system prompt every call? Cache it."
Tip: Add cache_control to an API request and repeated prefixes are read from cache at a fraction of the normal input price (0.1x on the standard multiplier). The default cache lives 5 minutes and refreshes on each hit; a 1-hour option exists.
Gotcha: Any change early in the prompt (tools, then system, then messages) breaks the cache for everything after it, so put the stable parts first.
2026-11-13-14 · T1 · Tip · Practical · "Don't need the answer right now? The Batches API costs half."
Tip: The Message Batches API runs requests asynchronously at 50% of the normal price. Most batches finish within an hour; good for evals, backfills and bulk classification.
Gotcha: Batches expire if not done within 24 hours; expired requests aren't billed, but you have to resubmit them.
Experience: NEEDS-BEN optional (a data job where overnight was fine)
2026-11-13-15 · T1 · Tip · Practical · "Still parsing JSON out of Claude's replies with regex? Stop."
Tip: Structured outputs constrain the response to your JSON schema, so the JSON is valid and the required fields are there. Set output_config.format with a json_schema; for tool calls, add "strict": true to the tool definition.
Gotcha: Not every schema feature is supported (no recursive schemas, no min/max numeric constraints, objects need additionalProperties: false); unsupported ones return a 400.