AI Agent Workflow Collection

Unattended work — "go ahead while I'm away", "run overnight", "proceed autonomously", a mandate for hours of solo work — and the return report after. Skip for ordinary interactive turns.

A wanted artifact with latent requirements — doc, plan, spec, schema, API shape, a look — where questions get "show me options" and requirements arrive as reactions. Also when a menu can't get a pick, and on "probe rounds". Skip when requirements are statable (interview) and for a new idea (superstack-inception).

Resuming or handing off — after a crash, session limit, compaction, or model handoff; picking up a HANDOFF or state file; closing a session with work in flight; the SessionStart continuity line. Also on "resume safely" or "write the handoff". Skip for a fresh session with nothing inherited.

Use when behavior contradicts expectation — a bug, an unexplained error, a failing test, output that looks wrong, a fix that didn't hold — or when you're about to retry the same category of fix a second time, or work keeps failing or stalling. Also when the user says "debug this", "why is this happening", or "still broken". Skip when the cause is already proven.

A mid-work technical fork — "which way should we go", a design/architecture/tooling choice about to lock — or the user ruling on how the thing is built (conduct rulings are superstack-doctrine's). Also on "record this decision". Skip for statute-governed choices, forks needing building (superstack-spike), and trivially reversible picks (a log line).

Dependencies need moving — an EOL or deprecation notice, a security advisory, "upgrade to vX", "how far behind are we". Also on "upgrade dependencies". Skip for new dependencies (superstack-decide), lockfile-only patch bumps, and updating the superstack plugin itself (the plugin manager's job).

A period report for a reader outside the session — a weekly update, client status, "what did we get done this month". Skip for end-of-task reports (superstack-ship), retros when a retro tool ships, and recalling past sessions for yourself (a memory tool).

A correction or standing ruling from the user — a rebuke of a working habit, "make this a rule", "always do X", "never do Y again". Mint the statute BEFORE giving the substantive reply. Also to review or amend standing law. Skip for one-off task instructions that don't outlive the turn.

Run a campaign — a build with milestone structure spanning many sessions — from an opened plan through audited milestones to closure, surviving compaction and cold starts. Load whenever the session-start line names an ACTIVE plan (required before build actions), when opening a plan, or on "continue the plan". Skip for work a task file covers.

An artifact with a face — a UI, dashboard, chart, game, rendered document, CLI output a human reads — whenever one is built or changed, before claiming done; and before trusting any new check or harness. Also on "look at it" or "screenshot QA". Skip for pure logic changes proven by existing tests.

A new idea with no spec — a product, tool, feature, or app arriving as "I have an idea", "let's build", "create/make me X". Also on "shape this idea". Skip for bug fixes, specced tasks, changes to a running codebase (superstack-scope) — and "new skill/module/pack" ideas (superstack-smith).

A live incident — production down or degraded, users affected, data at risk, a deploy gone wrong — however phrased ("why is this broken" included); mitigation outranks diagnosis while harm is ongoing. Skip for dev-time bugs with nothing live behind them (superstack-debug).

A schema or data change with real data behind it — DDL on a live table, a backfill, retyping/renaming columns, moving data between stores. Skip for empty schemas and rebuildable dev databases; code/dependency moves (superstack-deps); a live migration harming data (superstack-incident).

Going public — a push, PR creation or merge, a package publish, a release, an infra apply; run before any of them, and when the outward gate bounces a publish command. Writes the sweep receipt the gate checks. Skip for purely local work.

Parked work — "park this", "add it to the backlog", "someday", a shaped idea inception parks, a deferral with a revisit trigger; read the queue first when starting fresh work. Also on "what's parked". Skip for work starting now.

Use before trusting or shipping an answer, design, plan, analysis, or risky code change — when you are about to commit to it. Also when the user asks for a review, a red-team, a second opinion, or to check work adversarially. For pure code diffs, prefer a dedicated code-review command when the environment has one; this runner covers every other artifact and needs nothing installed.

Use at the start of a non-trivial task, or when the goal is fuzzy, the request is solution-shaped, or you cannot yet state in one sentence what "correct" will be checked against. Skip for trivial edits.

Use when finishing a task — writing the final report or status, shipping a change, or handing off — to produce a calibrated done-claim and the docs that make the work reproducible by someone else.

A new skill or pack for a superstack-family plugin — "new skill", "add a module", "build a pack for X" — or a recurring moment applying for admission. Skip for editing an existing skill (just edit, red-first near hooks) and product ideas that aren't skills (superstack-inception).

A feasibility question, not a delivery — "can X even do Y", "is this viable", "how expensive would Z be" — where throwaway code is the cheapest probe. Skip when a read-only probe answers, when reasoning settles it (superstack-decide), and for work meant to be kept (superstack-scope).

Use when the user asks for the superstack status, overlay health, the calibration record, or what the method has been tracking in this workspace. Read-only report; changes nothing.

The human wants to understand — "walk me through this", "explain at my level", "teach me" — or a handover where the maintainer is the human. Skip for one-line answers and stranger-facing docs (a documentation task, not teaching).

An outcome claim shipping — "this should increase signups", "this saves the team time" — recorded as a dated, falsifiable prediction; settled when the SessionStart line says due. Skip for work with no outcome claim.

Use when about to claim work is complete, fixed, passing, correct, or done — before committing, opening a PR, or moving to the next task. Also when a result looks good and you're tempted to move on, or a green signal came back suspiciously clean. If the environment ships a dedicated end-to-end verify skill for exercising code changes, prefer it for driving the change; this runner owns claim calibration (the evidence ledger) and non-code artifacts.