Ship 30 for 30: week 3 drafts (Days 17+, from Mon Oct 26)
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. Week 2 (Days 10-16) lives in [[2026-10-19-week2-drafts]]. Days 17+ follow the topic-bank order.
| Day | Date | Headline | Ship 30 rule | Words |
|---|---|---|---|---|
| 17 | Mon Oct 26 | 3 checks an agent should run before it says "blocked on you" | Subheads every ~100 words, mirroring the headline | 270 |
| 18 | Tue Oct 27 | Never let an agent grade its own homework | 1/3/1 rhythm under each subhead | 245 |
| 19 | Wed Oct 28 | Write the test before you let the agent build | Short-form: 1 line / 3 bullets / 1 line, Screenshot Test | 93 |
Day 17, Mon Oct 26
Rule: Subheads every ~100 words (halves or thirds) that mirror the headline promise exactly ("3 checks" → "Check #1/#2/#3"); plain-text subheads because LinkedIn has no bold (live session: Data-Driven Niche Selection). Source: topic-bank-v1 Day 17 row; [[2026-10-02-post2-agent-didnt-crash-v1|post2]] §5 "Verify the blocker at its source"; ~/.claude/state/ray-handoff.md 2026-07-31 correction trail and the Codex-critic DONE block (login 07-28, auth.json 08:00); memory feedback_verify_blockers_against_source_not_notes.md (the three checks: check the state, a note is a claim with a date, write uncheckable status as unverified with its date). Left out: the 2026-08-27 recurrence, because it happened on client work. Critic: fresh sw-critic r1 ITERATE (7 fixes: headline actor matched to body, "checks" vs "sweeps" disambiguated, ', not X' subhead cut, Check #2 logic fixed (the note was newer than the login, so date alone misleads), passive cut, caveat label cut); applied by parent, one narrowed: the suggested 'plus one test call' was not added because it was not part of what happened.
3 checks an agent should run before it says "blocked on you"
In late July, my AI agent told me four times in two days that a task was blocked on me. [CHECK: ray-handoff.md 2026-07-31 correction trail: 07-30 twice, 07-31 03:02, 07-31 morning brief] It needed me to log in to a second AI model it uses as a reviewer. I had logged in on July 28, the morning the task started. [BEN: confirm this framing is yours]
Nothing crashed. That evening, about 14 hours after my login, one of the agent's own task-board sweeps wrote a note: "needs founder login." [BEN: confirm; add why it wrote the note if you know] Every later sweep quoted the note.
Now my agent runs three checks before it tells me I'm the blocker. [BEN: confirm; this is the standing rule you gave it after you pushed back on 07-31]
Check #1: Look at the source
My login left a file with a timestamp: 8:00 that morning. One look at that file would have shown the login was done. [CHECK: Codex-critic DONE block, auth.json 08:00] The same goes for a deploy, a ticket or a pull request. If the blocker is checkable, check its state.
Check #2: Ask when the source last confirmed it
A note is a claim with a date. This note was about 14 hours newer than my login, so it looked current. No sweep ever checked it against the login file.
Check #3: If you can't check it, say so
Some sources sit behind access the agent doesn't have. Then the agent reports the status as unverified, with its date: "blocked, per a July 28 note, not rechecked." That tells me to look before I act.
These checks still trust the source. If the ticket itself is out of date, the agent will report it faithfully.
When your agent says it's waiting on you, how do you know it's true?
Day 18, Tue Oct 27
Rule: 1/3/1 rhythm under each subhead: one-sentence opener, a three-sentence middle, one-sentence close (variants 1/4/1, 1/2/5/2/1); subheads kept from Day 17 and mirror the headline's "grade" ("grade itself" / "who grades it" / "what the grader must read"), plain text because LinkedIn has no bold. Source: topic-bank-v1 Day 18 row; [[2026-05-19-verification-as-independent-worker-pattern]] (reviewer gets artifact + spec only, no producer framing; PASS/ITERATE/SCRAP; no auto-rewrite; escalate to founder when fixes keep failing); memory feedback_fresh_eyes_subagent_for_own_artifacts.md (2026-05-07 video defects and the founder's "fresh eyes" instruction); memory feedback_workflow_agent_output_integrity.md mode 5 (2026-07-10 false "verified against primary text" stamp, caught only by a zero-context reviewer reading the actual regulation). Distinct from Day 5 (what to do with findings) and Day 13 (self-scored data): this one is why self-review fails and what the reviewer must be given. Round cap left unnumbered on purpose: Day 5 said three rounds (studio gate), the SOP says two. Left out: the brand and product the video was for; the regulation's citation; the 2026-09-02 auth-brief example. Critic: parent-run content-critic ITERATE (1/3/2 rhythm break, reviewers checked the builder's file not notes, missing catch, "AI" unexpanded, intent claim) → 5 fixes applied by parent 2026-10-27. Round 2 (fresh sw-critic, 2026-10-27 ~16:35 ET): ITERATE, 5 exact edits (invented cause, inflated "reads" → "has to read", "nothing else" contradiction, missing [BEN] at escalation line); applied verbatim; critic stated PASS once applied. Fix 6 (grader→reviewer) declined to keep the headline mirror.
Never let an agent grade its own homework
In May, my agent made a short video and checked its own frames. It called the layout clean. I looked at the same frames and found two defects: a badge cut off at the right edge and a word wrapped onto a second line. [CHECK: memory feedback_fresh_eyes_subagent_for_own_artifacts.md, 2026-05-07 03:14 ET: badge clipping past the right edge at t=12s, "FAIL" wrapping to a second line in the outro at t=26s] [BEN: confirm this framing is yours and that you're fine saying you caught it]
Why an agent can't grade itself
The agent that built the work also checked it.
It read its own frames too generously. It said the cut-off badge "fits with margin." It called the closing frame with the wrapped word "clean." [CHECK: same memory file quotes the agent: "fits with margin" (t=12s) and the outro called "clean" (t=26s)]
I called it confirmation bias. [BEN: your 2026-05-07 words were "A bit crazy that you show confirmation bias."]
Who grades it instead
That day I told my agent to hand every review to a second agent with fresh eyes. [BEN: confirm; the memory file quotes you on 2026-05-07: "delegate to subagents with 'fresh eyes' to avoid confirmation bias of your own work"]
The reviewer gets the finished work and the spec. It never sees the builder's reasoning or the builder's "looks good." It returns pass, iterate or scrap, with each defect and a fix.
If the builder's fixes keep failing, the work comes to me. [BEN: confirm escalations reach you in practice] [CHECK: SOP 2026-05-19 "Max 2 verify-iterate cycles", then escalate to founder; Day 5 said three rounds, so no number here]
What the grader must read
A fresh reviewer can still inherit a bad source.
In July, one of my agents described a federal regulation wrong and stamped its own work "verified against primary text." Two independent reviewers passed it because both checked the file the builder wrote. A third reviewer with no context read the regulation itself and caught the error. [CHECK: memory feedback_workflow_agent_output_integrity.md mode 5, 2026-07-10: 16 CFR 255.6 defined as intermediary liability (real 255.6 = child-directed endorsements); a skill author and two independent critics propagated it, each checking the same intermediate file; a zero-context reviewer reading the actual regulation caught it] [BEN: confirm you're fine sharing this one]
Now at least one reviewer in every chain has to read the original source. [BEN: confirm this is your standing rule; memory rule (c): "at least one gate per chain must go to the primary"]
Who grades your agent's homework, and what do they read?
Day 19, Wed Oct 28
Rule: Short-form post (X / LinkedIn short post): 1 line / 3-5 bullets / 1 line, and it must pass the Screenshot Test (specific reader, clear angle, skimmable, worth a screenshot). The X cut is over 280 characters; trim the third bullet for X if needed. Source: topic-bank-v1 Day 19 row; memory feedback_plan_tests_implementation_order.md (founder 2026-07-25: "writing something and then writing the tests to pass. That gives us false confidence and a pat on the back"; "document our rationale for what needed to be changed on the fly"; "If an assertion cannot fail against the current implementation, it is not a test, it is an observation"). Topic-bank bullet "review gets cheaper" dropped: no source for it. Left out: the 7-skill rewrite that prompted the rule. Critic: fresh general-purpose r1 ITERATE (3 fixes: "spec" vs "plan" one-term rule, line 2 narrowed to the agent testing its own code (nemesis), over-resolved closing one-liner replaced with a reader action); all applied, the closing fix narrowed from the critic's "check that each test failed at least once" to the founder's own can-it-fail test, so nothing new is invented.
Write the test before you let the agent build
A test written after the code is a pat on the back. When the agent writes tests after its own code, they pass because they were written to pass. [BEN: echoes your 2026-07-25 words, "That gives us false confidence and a pat on the back." Confirm]
- Write each test from the plan, before the agent writes any code.
- The agent then builds toward a target it didn't set.
- If the plan changes mid-build, update the plan, then the tests, then the code. Write down why. [BEN: confirm this is your rule; 2026-07-25: "document our rationale for what needed to be changed on the fly"]
Before you merge, ask of each test: could it fail against this code?