The Pattern in One Paragraph
The previous four posts describe one pattern from different sides. A program starts from a vague sentence and becomes a small process with intake, review, ownership, and a cadence (post two). A recurring check keeps it moving, and a daily report covers the gaps between checks (post one and post three). Agents that support the program sit on a ladder from report-only to auto-execute, and each action earns its place on the ladder (post four).
Start With the Sentence
The hardest step in each program was the first one. A vague need has no structure to work with, and jumping to a solution skips the work of understanding the problem. Claude contributes most at this stage by asking questions. Each question forces a decision I would otherwise defer, and the answers become the program's design.
Constraints Come First
Setting limits before building made experimentation safer. A report-only agent can run every day because its failure is bounded. That bound let me try ideas quickly, since a wrong idea cost a minute of reading and nothing else. Later agents inherited the habit: each one starts at the lowest tier that lets it be useful.
Small Formats Grow
One report in one format grew into a routine because each addition answered a check I had done by hand twice. Growth driven by repeated need has produced reports that stay short. Growth driven by imagination tends to produce reports that get skimmed.
Ownership Beats Tooling
The programs that stayed healthy shared a trait: each open item had a person and a next date. The tools varied. When a program stalled, the fix was almost always to name an owner or set a date.
Definitions Need Owners
Two of the four report groups depend on definitions: what counts as quiet and what counts as finished. Those definitions come from how a team uses its tracker, and they drift as workflows change. Someone has to own them and review them on a schedule, the same way someone owns any other part of the program.
How Claude and I Work Together
The division of labor stayed consistent across every program. I supply context about the organization and make each decision. Claude asks questions, drafts artifacts, and tests designs against hypothetical cases. Claude's drafts give me something concrete to react to, and my reactions steer the next draft. Both sides contribute to every draft, and the results came from the exchange.
Limits of the Pattern
An overseer reports recorded state. Work that never reaches the tracker stays invisible to it. The report cannot decide priorities, resolve disagreements between people, or repair a tracker with no working conventions. Those tasks stay with people. The pattern reduces the work of noticing and leaves the work of deciding exactly where it was.
Open Questions
Several questions remain open. How often should definitions of quiet and finished get reviewed? When does a report section stop earning its place? Which parts of the pattern transfer to a team that works differently from mine? Does the independent reviewer approach hold up as agents take on more kinds of work? I expect answers to come from running more programs and comparing notes with readers.
Where I Go Next
I am considering three directions. First, applying the same structure to more programs, to see which parts transfer unchanged. Second, reviewing the definitions on a schedule of their own. Third, refining the skill template as readers try it and report what they needed to change.
Why I Write This Down
These posts describe structure and habits. Each reader will adapt them to an organization I have never seen, with different trackers, different teams, and different constraints. I share the pattern in the hope that people who run programs of their own will find a piece they can use, and I welcome corrections from anyone who has tried something different.
For Readers Starting Out
A practical sequence for anyone who wants to try this: pick one program, write the one-page description, add a report-only overseer, and run it by hand until the definitions settle. Put the report on a schedule after that. Add a section only after repeating a manual check twice. Every step in that sequence is small, and each one can stop the work from going stale.
A first month might look like this. In the first week, write the one-page description for a single program and name its owners. In the second week, run a report by hand and adjust the definitions. In the third week, put the report on a schedule. In the fourth week, hold the first review that uses the report as its starting material, and record what the review decided.