The field-notes gated pipeline (this very L2 assess stage) already runs on Claude Code CCR triggers, so a Claude Code skills framework like superpowers applies directly to existing infrastructure. No new services, languages, or hosting are required - it is a prompt-organization pattern layered on tooling already in use.
The operator favors practical, budget-conscious, unattended systems over experimental architecture work. SDD is a methodology and prompt-restructuring exercise, not a new capability or user-facing feature, so it only fits the 'runs unattended' value if it demonstrably improves debuggability of stages that are currently hard to maintain - which has not been shown yet.
Implementation Options
01Pilot: decompose L2 Assess into skills
field-notes3-5 hourslow risk
Extract this assess stage's monolithic prompt into 2-3 small Claude Code skill files (e.g. queue-check, assessment-builder, artifact-writer) under field-notes/.claude/skills/, invoked by a thin orchestrator. Run it against the current assess_pending backlog to compare output quality and debuggability against the existing monolithic version before touching L3/L4.
+Tests the pattern on the lowest-risk, most self-contained stage
+Improves debuggability of the stage most likely to change over time
+Reusable skill files could seed similar decomposition in bud's or codec's pipelines
-Refactors a stage that currently works, with no reported failures to fix
-Adds file/directory structure and indirection for a solo operator to maintain
-Time spent here is not spent on new capability
02Full SDD refactor across L2-L4
field-notes2-3 daysmedium risk
Apply the same decomposition to the plan and execute stages as well, building a shared skills library (scope-analyzer, plan-reader, file-editor, PR-creator) reused across all three gated stages.
+Consistent architecture across the whole pipeline
+Named methodology makes onboarding/documentation easier if the pipeline grows
-Touches the execute stage, which writes to git/PRs - a bad refactor there has real consequences for an unattended system
-Multi-day effort with no user-facing payoff
-Premature given L2/L3/L4 have not been shown to have reliability problems
03No action now, revisit on pain
field-notes0 hourslow risk
Skip implementation. Keep the current monolithic stage prompts. Reconsider SDD-style decomposition only if a specific stage becomes hard to debug or extend.
+Zero effort, zero risk to a working unattended pipeline
+Avoids abstraction before it is needed
-Forgoes the debuggability benefit until forced to deal with it under pressure
-Loses the shared vocabulary/pattern for future multi-agent work
None immediately - the current gated pipeline stages already run successfully. The concrete value, if pursued, is faster debugging and easier extension of individual pipeline stages (starting with this assess stage) via small, testable, named skill files instead of one large prompt per stage.
The field-notes pipeline is the template for any future multi-stage agent workflow at Understory Labs (bud's email pipeline, a future codec content pipeline). Naming and trying the SDD pattern once, on a low-risk stage, gives a reusable answer for how to structure the next unattended multi-agent system without committing to a full rewrite now.
03 · What Implementation Looks Like If pursued: create field-notes/.claude/skills/assess-queue-check, assess-build-artifact, and assess-write-advance as separate skill files, each with a narrow input/output contract. Replace this routine's single prompt with a short orchestrator that calls them in sequence. Nothing changes for the operator day-to-day - the L2 stage still produces the same JSON artifact and comment thread - but the underlying prompt is now three small, independently testable pieces instead of one long one.