End-of-week orchestrator

Summary: A skill that runs end-to-end on Friday afternoons, chaining together three workflows that used to run separately: a quarterly progress KPI report, a newsletter analytics digest, and a structured reflection across calendar, downloads, email, notes, and weekly wins. The orchestrator runs each stage in sequence, passes the outputs forward, and produces a single reflection file with everything linked. The user shows up to answer the reflection prompts and reviews the consolidated output.

What makes this automation-level: The user triggers the run once and the AI handles each defined stage in order without further direction. Each stage was previously a separate manual workflow with context-switching between them; chaining them flips the pattern from "user orchestrates AI tools" to "AI orchestrates the run, user shows up at decision points." Reflection prompts are still answered by the human, but the framing, sequencing, and synthesis are handled by the orchestrator.

Map

The recurring workflow is the weekly review on Friday — checking quarterly progress, reading newsletter performance, and reflecting on the week. Decomposed into: pull KPI data from calendar and project management, draft the progress report, query the newsletter analytics, draft the audience insights digest, run reflection prompts in a specific order (calendar week recap → downloads/desktop captures → email triage → notes/tasks → wins and brain dump), generate a final reflection file linking all outputs.

Almost every step except the reflection prompts themselves was a good AI candidate. The reflection prompts deliberately stayed human-driven — that's the point of the review — but the framing and sequencing got delegated.

Test

Initial version ran each stage as a separate skill invocation. Each one worked independently but the Friday review took an hour of context-switching: open one tool, run the report, copy results, open the next, repeat. The fix was orchestration — a skill that calls the sub-skills in order, passes outputs forward as context, and produces one final file.

The first orchestrated run cut the total time roughly in half, but the more important effect was that running it became reliable. A multi-step manual workflow gets skipped or shortened on a busy week; a one-trigger orchestrated workflow runs in full.

Integrate

Built as a top-level skill that calls three sub-skills in a fixed order, then walks the user through reflection prompts via a waterfall. The KPI report stage runs first because it produces the data the audience insights stage uses to compare metrics. The audience insights stage produces content opportunities that the reflection prompts can reference.

The reflection waterfall is structured so each prompt builds on the previous one — calendar before email before notes before wins — to surface patterns rather than collecting flat answers. The final output is a markdown reflection file with sections for each stage and links to the KPI report, the audience insights file, and any notes referenced during reflection.

Compound

Two compound mechanisms run in this workflow.

The reflection prompts themselves refine. After running the orchestrator for several weeks, certain prompts consistently surface useful insight ("what did you say no to this week?") while others mostly produce filler ("rate your energy 1-10"). The high-signal prompts get kept; the low-signal ones get rewritten or dropped. The order also gets tuned — putting calendar review before email triage turns out to surface different patterns than the reverse, because seeing what the week actually contained first changes how the inbox gets read.

The integration between stages compounds. Early versions ran each stage independently and the user did the cross-referencing manually. Over time, downstream stages got tuned to use upstream context — the audience insights stage now references this week's KPI movement when suggesting content angles, and the reflection prompts now reference both. Each cross-reference saves a manual step and makes the next run slightly tighter.

agent-design-principles — The orchestrator embodies the "context as expertise" principle: each stage carries forward only what the next stage needs, not the entire run history
compound-engineering — Both compound mechanisms (prompt refinement and stage-integration tightening) are concrete examples of continuous improvement through use
ai-productivity-traps — This workflow only works because the human still answers the reflection prompts; an over-automated version that auto-generates reflection answers would produce decorative output that nobody actually reads or acts on