feat(claude-ops or session-flow): loop-prompt authoring skill — generate and maintain standing loop-lane prompts

[kyle-sexton]

Category: improvement · Source: 2026-07-19 loop-lane operations — the babysit/triage/work lane prompts went v1→v4 in 24h via hand-editing; every iteration fixed a class of failure a skill could have prevented.

Problem

Standing prompts for /loop lanes are hand-authored per session. Hard-won structure — role statement (orchestrator vs implementer), cycle-bounded batch caps (a cap must bound a CYCLE, never terminate a self-paced loop), self-observation/dogfood rules, dedupe-first filing, error-handling STOP conditions, cycle status lines, week-directive parameterization — lives in one user's scratchpad files and gets re-invented or mis-worded (a mis-worded cap silently killed a lane overnight).

Proposal

A skill that authors and maintains loop-lane prompts: takes lane role + target skill + scope + directives as input, emits a standing prompt from a canonical template embedding the invariants above; supports versioned regeneration (diff what changed); warns when a rule belongs in the target skill instead of the prompt (absorption candidates, cf. #477/#478/#479).

Evidence base

.work/leave-week-priorities/loop-prompts/ v1→v4 evolution; #479 comment on the cycle-vs-session cap failure.

Placement (claude-ops vs session-flow vs autonomy) = design decision for triage/human.

---

Reference: Boris Cherny, "Steps of AI Adoption" (Google Doc)

[kyle-sexton]

This was generated by AI during triage.

Triage: routed needs-human + wayfind: design

Category: improvement · Priority: priority: high (week directive: the pipeline improving itself is this week's second priority).

Genuine open design space on two axes the author explicitly punted — placement and the authoring mechanism itself — so this is human-gated rather than defaulted. Below is a design-options sketch to make the human decision cheap.

Open decision 1 — placement (author-flagged: claude-ops vs session-flow vs autonomy)

Candidate homes surveyed among installed plugins:

Home	Fit	Against
autonomy (RECOMMENDED)	Owns governed autonomous agent operation — role-topology, trigger-dispatch, climbing the AI-adoption ladder. A standing /loop lane is a governed autonomous trigger; a skill that authors/maintains its operating contract is squarely this plugin's domain.	Currently nascent (only a setup skill) — this would be its first operational skill.
session-flow	Closest existing sibling is orchestrate (arms a session/worker with proactive-orchestration imperatives); loop-prompt authoring is adjacent to that arming step.	session-flow is scoped to this session's lifecycle; standing loop lanes outlive a session, which is more autonomy's frame.
claude-ops	Ops toolkit (observability, plugins, changelog).	Its skills read/operate product telemetry & fleet, not author agent operating contracts — weakest conceptual fit.

Recommendation: autonomy — the loop lane is a governed autonomous trigger and this skill authors its contract; being autonomy's first operational skill is a feature, not a blocker.

Open decision 2 — authoring mechanism (genuinely open, needs design)

The proposal names several mechanisms that each need a design pass before implementation:

• Canonical template embedding the hard-won invariants (role statement orchestrator-vs-implementer; cycle-bounded batch caps — a cap must bound a CYCLE, never terminate a self-paced loop, the mis-wording that silently killed a lane overnight; self-observation/dogfood rules; dedupe-first filing; error-handling STOP conditions; cycle status lines; week-directive parameterization). Template format + parameterization surface is undesigned.
• Versioned regeneration ("diff what changed" v1→v4) — needs a storage/versioning model.
• Absorption-warning heuristic — warn when a rule belongs in the target skill instead of the prompt (cf. source-control(babysit-prs): absorb v4 loop-prompt operating rules into the skill #477/work-items(triage): absorb v4 loop-prompt routing rules into the skill #478/work-items(work): absorb v4 loop-prompt execution rules into the skill #479). Detection rule is undesigned.

Recommended human path

1. Pick the home (recommended autonomy).
2. Run /planning:design (or /planning:wayfind) on the authoring mechanism — the template schema, versioning model, and absorption-warning heuristic are the design deliverables.
3. Then it becomes an implementable brief.

Evidence

.work/leave-week-priorities/loop-prompts/ v1→v4 evolution; #479's comment on the cycle-vs-session cap failure. Sibling absorption issues: #477 (babysit), #478 (triage), #479 (work) — this skill manages the prompt-level residue those three leave behind, so it is complementary, not competing.

[kyle-sexton]

Cross-reference from the #481 design-confirmation session (2026-07-19): confirmed this stays a separate skill from the goal-condition crafter — loop prompts have no evaluator or character-limit semantics, and the discovery concerns differ (docs/PLUGIN-PHILOSOPHY.md split rule). A universal prompt-authoring generalization was considered and deferred; trigger = observed duplication after both skills ship. Details: #481 (comment)