01-projects/personal-brand/ship30

week1 drafts

2026-10-07·draft·status: awaiting-founder
ship30linkedin

Ship 30 for 30: week 1 drafts (Days 1-7, Oct 10-16)

Ray drafts, Ben edits and posts. Each piece is an Atomic Essay, 300 words or fewer. Word counts include the headline and exclude the [BEN] and [CHECK] notes. No hashtags. Plan: Ray proposed it on 2026-10-07; the founder asked for week 1 drafts at 21:04 ET ("Drafting out my first week would be a great head start"). The structure still awaits his read. Day 1 announces through the session's four questions, Days 2-5 take one question each, Days 6-7 keep the topic-bank rows.

Day Date Headline Ship 30 rule Words
1 Sat Oct 10 4 questions engineers asked me about building software with agents Announcement essay (value-first variant) 246
2 Sun Oct 11 The spec your agent writes is disposable. Keep the tests and the why. 2-Year Test 257
3 Mon Oct 12 Your agent forgets everything between sessions. Mine reads a knowledge base. For Who / So That 251
4 Tue Oct 13 Is spec-driven development overkill for a bug fix? Shrink the spec, don't skip it. Personal credibility 260
5 Wed Oct 14 When an agent reviewer finds something outside the spec, send it back with a test 4A, Actionable cut 248
6 Thu Oct 15 4 reasons the one job agents can't do is your whole business Endless Idea Generator, Reasons container 249
7 Fri Oct 16 The 4 steps I used to score 2,500 small businesses on 5 jobs Lean Writing, Steps container 252
8 Sat Oct 17 3 reasons admin is the first job to hand to an AI agent Headline: specific topic + proven approach, Reasons container 269
9 Sun Oct 18 How small business owners can hand outreach to agents without spending their trust Headline + explicit audience 282

Day 1, Sat Oct 10

Rule: announcement essay, value-first (Ray-proposed 2026-10-07, awaiting his read). Source: [[2026-10-07-locked-topics-and-question-bank|locked-topics question bank]].

4 questions engineers asked me about building software with agents

About 30 engineers asked me how to build software with agents. Four of their questions, with a short answer to each.

The context: this week I taught them an advanced Claude Code session on the agentic software development lifecycle. It's spec-driven: before the agent writes code, it writes three files: an intent file (what the change is for), a plan file (how), and a test file (how you'll know it worked).

  1. Are those documents thrown away once the work ships? Yes. They're scaffolding for one change. The tests are what stay. [BEN: confirm the second and third sentences; your answer in the room was "yes"]

  2. Isn't there long-running context? How do you capture tacit knowledge? A knowledge base the agent reads before it starts. For an enterprise, how to manage that context is still an open question.

  3. Is spec-driven development overkill for a bug fix? This one and the next didn't get a full answer in the room. Here's my answer now. Scale the spec to the size of the change. A one-line brief plus a failing test is a spec. [BEN: confirm]

  4. When an agent reviewer flags something outside the spec, do you open a new task, send it back to the builder, or add a test? Send it back to the builder with a test that catches it, and cap the rounds so the loop can't spin. [BEN: confirm]

I'm writing about making agents useful in the enterprise every day for the next 30 days. These four questions are where I'm starting.

Day 2, Sun Oct 11

Rule: 2-Year Test (teach the version of you from 2 years ago). Question 1.

The spec your agent writes is disposable. Keep the tests and the why.

At a session I taught on Wednesday, someone asked whether the intent, plan and test documents an agent writes get thrown away after the work ships.

Yes.

Two years ago I would have argued the other way. I'd have wanted every spec filed and kept, because good documentation was slow to write and rare to find. [BEN: confirm this is true of you two years ago, or swap in what you actually believed]

Now an agent writes a plan in a minute. Cheap documents change what's worth keeping. [BEN: confirm]

I run an always-on agent at home. Every job it hands to a sub-agent starts with a written brief. The brief is single-use. Once the work is merged, nobody reads it again.

Two things survive:

Everything else in the brief was scaffolding. It held the work up while it was built.

The mistake is keeping the scaffolding and losing the decision. A folder of old plan files looks like documentation. It mostly records what someone intended once, not what the code does now. [BEN: confirm]

If you do spec-driven development with agents, try this at merge time:

What does your team keep after an agent-built change ships?

Day 3, Mon Oct 12

Rule: For Who / So That at the top of the page (pre-draft note, not posted). Question 2. TOPIC: long-running context for agents. FOR WHO: engineering and data leads rolling out coding agents. SO THAT: their agents stop relearning the same lesson every session.

Your agent forgets everything between sessions. Mine reads a knowledge base.

The second question from last week's session: if the spec documents get thrown away, where does long-running context live? And how do you capture tacit knowledge?

My answer was a knowledge base.

Here's what that looks like for the always-on agent I run at home:

The last one is how tacit knowledge gets captured. Not by interviewing anyone. By writing down the correction the moment it happens. [BEN: confirm]

One example. My agent once told me four times across two days that a task was blocked on my login. I had logged in days earlier. It was quoting an old note instead of checking. That correction is now a standing rule: check a blocker at its source, not in a note.

For one person, this works. For an enterprise, I don't think anyone has it solved yet. [BEN: confirm]

The open questions are about ownership, not storage: [BEN: confirm]

Those sound like data governance questions. I think they are. [BEN: confirm]

Where does your team's tacit knowledge live today?

Day 4, Tue Oct 13

Rule: personal credibility (name the experience that qualifies you). Question 3. Credibility: his own always-on agent.

Is spec-driven development overkill for a bug fix? Shrink the spec, don't skip it.

The third question from last week's session: spec-driven development makes sense for features. But for a bug fix?

Fair question. Writing an intent file, a plan file and a test file for a one-line change feels like filing paperwork to fix a typo.

My take comes from the agent I run at home on a Mac mini. It runs scheduled jobs every day, and sub-agents do most of the building. I've handed it plenty of fixes, big and small.

What I've landed on: scale the ceremony to the size of the change. Don't drop the spec. Shrink it. [BEN: confirm]

For a bug fix, the whole spec is two things:

That's intent and test in their smallest form. The plan is usually obvious once the test fails for the right reason. [BEN: confirm]

An example from my own setup. My daily research scan timed out two days in a row. It didn't crash. It sat there until the scheduler's time limit killed it. That bug didn't need a plan file. It needed one line ("the scan finishes inside its time limit") and a run that proves it. [BEN: confirm this framing; the real fix was moving the reading inline, away from background agents]

The small spec buys the same thing the big one does. The agent can't call the job done until the test passes. And the test stays after the brief is gone. [BEN: confirm]

Where I'd still write the full set: anything that touches more than one system, or anything you can't easily undo. [BEN: confirm]

Where do you draw the line?

Day 5, Wed Oct 14

Rule: 4A (run one topic through all four A's, ship one). Shipping the Actionable cut. Question 4.

When an agent reviewer finds something outside the spec, send it back with a test

The fourth question from last week's session: agent code reviews catch things outside the spec. Do you open a new task, send the work back to the builder, or add a test that catches it earlier?

Here's what I do with the agent I run at home.

  1. Review with fresh eyes. The reviewer is a separate agent. It gets the finished work and the rules, and none of the builder's reasoning. A builder can't see past its own brief. [BEN: confirm]

  2. Send each finding back to the builder with a test. Not a note. A test that fails until the defect is gone.

  3. Cap the loop at three rounds. If it still fails after round three, the loop stops and escalates to me.

A recent example. A builder agent changed a catalog sync I run. The fresh reviewer flagged defects the build brief never named, including:

Neither was in the spec. Each went back to the builder with a test.

Source (not for posting): fresh-critic reports on the catalog-sync PR, 2026-10-07 (rollback plan left the catalog test red; production sync had no merged-code guard). Both fixed with tests; a third critic round passed.

So my answer to the three options is all three, in order. [BEN: confirm]

How does your team handle review findings nobody asked for?

Spare 4A cuts on the same topic (not for posting this week):

Spare ideas for Day 6 (not for posting): 3 businesses, one shape; why admin tops 11 of 12 types; score your own five jobs; why delivery is lowest in 9 of 12 types; automation is the floor, not an edge.

Day 6, Thu Oct 15

Rule: Endless Idea Generator, "Reasons" container. Source: [[2026-10-04-fifth-job-01-four-of-five-v2]] (figures and foil gate from that draft's receipts).

4 reasons the one job agents can't do is your whole business

Most small businesses do five jobs: find customers, sell, deliver, support, and admin.

I had a model score 2,500 public businesses from Starter Story on how much of each job an artificial intelligence (AI) agent can carry today, from 0 to 3. The median business scored four of its five jobs at 2 or more, which I call agent-shaped.

In nine of twelve business types, the job left over is delivery. Here's why it's the whole business. [BEN: confirm the four reasons]

  1. Customers don't pay by the job. Nobody buys from a furniture maker because the invoices go out on time. They buy the table.

  2. The zero is a different kind of work. Frailey Woodworks, a custom furniture shop, scores 2 on four jobs and 0 on deliver. That 0 isn't a hard task scored low. It's hands and a workshop.

  3. It holds at scale. Deliver scores 0 for 139 of 142 local service businesses in the data. Admin scores 2 or more for 139 of the same 142.

  4. Automation becomes the floor. If every pet sitter hands admin to an agent, admin stops being an edge. The fifth job is the only place left to compete.

So the question isn't "can agents run my business?" It's "which jobs should agents carry, so people get more hours for the fifth?"

One caveat: a model scored itself from 300-character public blurbs. Treat it as a starting frame, not a measurement.

What's your fifth job, in one sentence?

Day 7, Fri Oct 16

Rule: Lean Writing, "Steps" container, ship small. Source: same draft's Receipts (method) and bear case.

The 4 steps I used to score 2,500 small businesses on 5 jobs

I wanted to know how much of a small business an artificial intelligence (AI) agent can carry today. Here's how I scored 2,500 of them, weak step included.

  1. Collect public stories. Starter Story publishes founder interviews with monthly revenue and team size in the header. A script deduplicated 3,261 stories into 2,566 businesses and dropped 66 with suspect revenue figures. That left 2,500.

  2. Sort them. One fixed prompt read a 300-character public description of each business and assigned one of twelve types, from software to local services.

  3. Score five jobs from 0 to 3. Find, sell, deliver, support, admin. A 0 means an agent can't do it, 1 means it assists, 2 means it does most of the job, 3 means it runs it.

  4. Spot-check. A second pass re-read 30 of the blurbs. The type matched on 28, and every job score was within one point.

Step 4 is where I'd push back on my own work. The check wasn't independent. The agent that wrote the prompt also ran it, against the same short blurbs. Nobody read the full stories.

And the whole thing rests on a model grading its own ability from a paragraph. A blurb about a woodworking shop doesn't say how custom quotes get made.

So I treat the scores as a map of where to look, not a measurement. [BEN: confirm]

If you were checking this by hand, which step would you redo first?

Day 8, Sat Oct 17

Rule: Headline = specific topic + proven approach (clear, not clever), "Reasons" container. Source: [[agentify-3000-series]] spine #2; [[archetypes]] (figures re-derived from data/businesses-scored.csv on 2026-10-17). Critic: ITERATE x2, fixes applied.

3 reasons admin is the first job to hand to an AI agent

If you run a small business and want to try an artificial intelligence (AI) agent, start with admin: bookkeeping, invoicing, scheduling, reporting.

I had an AI model score 2,500 public businesses from Starter Story, a site of founder interviews, on five jobs: find customers, sell, deliver, support, and admin. [BEN: confirm you ran this scoring] Each job got a score from 0 to 3 for how much of it an agent can do today. The model scored each business from a 300-character public blurb, so treat the scores as a rough first sort.

  1. The model scored it highest. Admin was the top job in 11 of 12 business types, and no type averaged below 1.9 of 3. Across all types, admin averaged 2.2. The other four jobs averaged 1.4 to 1.9. [CHECK: re-derived 2026-10-17 from businesses-scored.csv, 2,463 typed clean rows: admin 2.17, sell 1.89, find 1.88, support 1.83, deliver 1.39; lowest admin type is ops at 1.95]

  2. Scheduling and bookkeeping mistakes are often cheap and easy to see. A double-booked call or a miscategorized receipt shows up at the next review, and you can undo it. Compare that with an agent misquoting your biggest customer.

  3. It's where you learn to check an agent's work. Admin leaves a record: a calendar, a ledger, a report. You can compare it with what happened. Those are the habits you'll need when you hand off riskier jobs. [BEN: does this match how you learned to trust your own agent?]

One exception: admin is also where the money moves. A wrong invoice or a duplicate payment costs trust fast. So the agent drafts, and a person approves anything that moves money or leaves the building. My own agent drafts my email. I press send. [BEN: confirm]

Which admin task would you hand off first?

Day 9, Sun Oct 18

Rule: Headline + explicit audience. Source: [[agentify-3000-series]] spine #3 (restated as "agents find; a human vouches" per its CHECK note); [[2026-10-04-kutay-gtm-in-git]] (self-reported, sponsored). Critic: ITERATE x2, fixes applied (unsourced "public"/"mostly online" sample claim left out pending a check).

How small business owners can hand outreach to agents without spending their trust

If you run a small business, let an artificial intelligence (AI) agent find customers and draft the outreach. Keep a person on anything that carries your name.

I had an AI model score 2,500 businesses from Starter Story, a site of founder interviews, on five jobs: find customers, sell, deliver, support, and admin. [BEN: confirm you ran this scoring] Each job got a score from 0 to 3. Finding customers averaged 1.9, and a 2 means the agent does most of the job. The model scored each business from a short public blurb, so treat that as a rough first sort. [CHECK: 1.88 across 2,463 typed clean rows, re-derived 2026-10-17 from businesses-scored.csv (Day 8 note); saas 1.95, consumer_app 1.93 per agentify-3000-series spine #3]

The research, the prospect lists, the first draft of an email, the search engine optimization (SEO) page: that work is agent-shaped. Some larger sales teams say they already run agents for research and audience building. [CHECK: John Kutay, "GTM in git", 06-reference/2026-10-04-kutay-gtm-in-git.md; Rippling, self-reported, promotes its own GrowthOS]

Being vouched for is not. Many small businesses grow by referral: someone the buyer trusts said it was good. An agent can't lend you that trust. It can only spend it.

So split the job:

  1. The agent finds. It researches prospects, builds lists, and tests which message gets replies.
  2. The agent drafts. Emails, posts, pages.
  3. A person reads anything with your name on it before it goes out.
  4. A person closes. The call, the referral ask, the apology after a bad week.

My own agent drafts my emails. I decide what goes out under my name. [BEN: confirm]

One warning. If every competitor's agent cold-emails the same inbox, that channel stops working. Volume is cheap now, so it stops being an advantage.

Which outreach task would you hand off first, and which would you never hand off?