The Everyday Track Sat at 18% While the Schedule Wanted 40%

Oct 20, 2026 · 3 min read

Cover image for: The Everyday Track Sat at 18% While the Schedule Wanted 40%

Roughly 131 of our 721 published posts were Everyday posts, about 18%, and the schedule wanted about 40%.

I was reviewing the blog and asked how we could fill out more Everyday posts. A read-only sub-agent counted them. The reader_path frontmatter field decides the track, and build-log is the default when it’s missing. Everyday posts carry everyday-ai. The rhythm is 3 Build and 2 Everyday per week, with Tuesday and Thursday as Everyday days, which is where the 40% comes from.

The numbers we acted on

Seeds: 517 total, 51 Everyday and 466 technical.
Everyday seeds: 5 open, 43 drafted, 3 held. Technical: 82 open, 379 drafted.
Drip stock from the blog-funnel command: build 1 of 6, everyday 1 of 4.
The funnel’s own line: “everyday stock 1/4 with no candidate left to stamp and 0 Everyday seeds ready to draft”.

The daily sweep, seeds_sweep.py, reads the work log, JIRA and BizOps. Those sources are mostly agent and infra work, and personal lines get stripped to keep the blog pseudonymous. The only real Everyday supply so far was a manual batch of 15 seeds on 10/1/26.

The diagnosis I got back was: “The problem is supply, not scoring. As far as the agent checked, nothing penalizes Everyday drafts, though it didn’t fully read the blind-score rubric.”

I said “Ok go.” That was a yes to a diagnosis that had admitted, in the same paragraph, that the rubric wasn’t fully read.

What we built on the supply diagnosis

Two agents ran in parallel.

The first wrote 16 Everyday seeds (ids 537 to 552). All 16 passed the wrong_belief_problem gate against their real evidence bundles and validate, with nothing loosened. Open Everyday seeds went from 5 to 21, and the funnel flipped from 0 to “16 Everyday seeds ready to draft”.

The second changed the sweep. seeds_sweep.py now makes a second Haiku call per day after the main pass. It keeps the existing privacy.py filter, adds filter_personal_text before the model sees anything, forces rows to non-tech, and requires evidence to be a verbatim line from the context. --no-everyday turns it off. funnel.py now sends one notification when fewer than 4 Everyday seeds are open, and it re-arms once the count recovers. Twenty-six targeted tests passed.

On supply, it worked.

What the queue did

The drip stock still said everyday 1/4. Stamping is a separate step, so the refill agent hadn’t touched it. We kicked off a draft run, which runs the normal daily blog pipeline by hand with --ignore-gate (it skips only the off-hours and usage timing check) and --live.

It drafts at most 2 posts per run. After 30 minutes, 5 of 16 seeds had drafts.
On the second pass, 11 of 16 had drafts and the pipeline’s cap of 4 drafts per night had been hit.
Of the first pass: 1 post reached the queue’s review step (“website won’t load”), 2 failed the final quality checks (“splitting costs” and “page nobody uses”), and “email sends” was waiting on older gate results to refresh.
Two Everyday posts did get stamped into the drip, and they came from older drafts. The log couldn’t tell us whether the draft agent did it.

So I had spent a refill and a new sweep on supply, and the drip stock was still 1 of 4. The report I approved also had a table with 71 Everyday drafts and 33 Everyday posts in review, which are not seeds waiting for supply.

Two funnel lines

These are the two real funnel lines, before and after the refill:

everyday stock 1/4 with no candidate left to stamp and 0 Everyday seeds ready to draft
Everyday seeds open: 16 (minimum 4), fine

The funnel counts draftable seeds, which is why the second line says 16 while the agent’s own count said 21 open.

Gaps we found along the way

The sweeps for 10/19/26 and 10/14/26 stopped early on an existing check that blocks a collaborator’s last name, so neither pass produced seeds on those days.
The last sweep before the change was at 6:15 am and gave 6 seeds, all technical.
We filed a follow-up due at 7:00 am on 10/20/26 to check that the 6:15 am sweep produced Everyday seeds.

The alert I shipped watches that draftable count, which read 16 after the refill. The drip stock still read 1 of 4 after the refill, because stamping is a separate step.

AI Skills

Use this lesson with the AI assistant you already use

Everyday posts were about 18% of 721 published posts against a 40% target, and a diagnosis of weak supply led to a 16-seed refill and a new sweep. The seed count recovered, but the drip stock the schedule actually draws from stayed at 1 of 4.

Paste the prompt, share only the context needed to answer it, and treat the result as a draft for your review. Do not include confidential information or let an AI assistant make changes without your approval.

The prompt works in any AI assistant as it is. We are also building a /dxdev skill for visual reports and saved memory. It is not released yet: see what you can use today.

dxdev LESSON · paste into your AI coding agent
LESSON: Verify The Metric You Are Trying To Move, Not Just The Count That Your Fix Improves
SOURCE: dxdev.com/blog/channel-mix-target-starved-by-supply
WHAT HAPPENED: A read-only sub-agent counted the posts and concluded the problem was supply, while admitting it had not fully read the blind-score rubric. The author approved that diagnosis anyway. Two agents then wrote 16 new Everyday seeds and added a second daily sweep with a new alert. Draftable seeds rose from 0 to 16, and the 26 targeted tests passed. A manual draft run, however, hit the 4-drafts-per-night cap, and only 11 of 16 seeds had drafts after the second pass. One post reached review and two failed quality checks. The drip stock still read 1 of 4 because stamping is a separate step that nobody had touched. The alert also watches the draftable count, not the stock the schedule consumes.
THE RULE: Before approving a diagnosis, check what it did not examine, and measure success at the stage that feeds the target, not at the stage your fix improved. A healthy upstream count says nothing about a downstream bottleneck.
CHECK MY CODE, then report PASS or FAIL with file:line for each:
      1. Does every diagnosis you approve state which inputs it did not read, and have you resolved those gaps before spending work on the fix?
      2. Does your success metric and any alert measure the final stage that feeds the target (here, drip stock), rather than an upstream count (here, draftable seeds)?
      3. Have you listed each pipeline stage between supply and output, including caps and manual steps such as stamping and per-night draft limits, and confirmed each one can carry the new volume?
THEN PRINT: a table (check, PASS/FAIL, evidence, fix) + a verdict (applies / partially / OUT_OF_SCOPE / no) + the single most important next action.

Send us what your agent found →

Prefer to talk it through first? Start an assessment or talk to us.

Lesson: Verify The Metric You Are Trying To Move, Not Just The Count That Your Fix Improves

Choose your next level of detail

Want the practical version?

How AI Helps starts with the work problem, what improves, and where human judgment still matters.

See how AI helps Talk to us

The Everyday Track Sat at 18% While the Schedule Wanted 40%

The numbers we acted on

What we built on the supply diagnosis

What the queue did

Two funnel lines

Gaps we found along the way

Use this lesson with the AI assistant you already use

Related in Running the work

When Yesterday's Design Decision Kills Today's Work

Our scoreboard dropped to 44% while the bookkeeping loop went uncounted

Two Posts About 956 Certificates Sat in the Pipeline Together

Grouping Open Items by Creation Source Changed Our Total

Nothing on the Screen Said the Setting Was Stored

Want the practical version?

Get the next Follow the Build post