| name | intuitive-flow |
| description | Stable execution/change router after an approved plan, preflight contract, or
tiny concrete task. Refactor-shaped work delegates to intuitive-refactor, and
durable work runs through staged planning, review, GSD handoff,
implementation, cleanup, and verification while keeping plan ledgers and
active capsules compact by replacing stale state instead of appending history.
Also accept vague prompts as a compatibility route that names the upstream
planning stage instead of hiding planning inside execution.
|
Intuitive Flow
$intuitive-flow is the stable build/change entrypoint. Accept the trigger,
then choose the smallest executor that preserves source-of-truth clarity,
verification, and stop gates.
Default to less. Before recommending or implementing a new file, API, module,
workflow, test layer, worker, plan, or compatibility surface, check whether the
same user-visible outcome can be reached by narrowing scope, deleting stale
surface, merging duplicate behavior, reusing an existing route, or documenting
current truth. Add a new entity only when that smaller option cannot satisfy the
accepted behavior and proof gate.
This is the compact runtime entrypoint. Use the Read First table below to load
the narrow reference needed for the selected route.
Modes
| Mode | Use when | Output | Redirect when |
|---|
| Direct route | Status checks, diagnostics, or tiny concrete edits are bounded enough for the main session. | Short route note, focused proof, and concise closeout. | The task is really discovery, refactor scope, or preflight. |
| Planning route | The user has a fuzzy idea, draft plan, or wants agents to align before execution. | Upstream recommendation through reduce-entropy, planning loop, grill-batch, or preflight. | The plan is already approved and execution-ready. |
| Durable execution | An approved plan/preflight/phase needs implementation, verification, docs, commits, and closeout. | Route brief, staged execution, proof card, docs alignment, parked todos. | The task is a known refactor seam or changed-code cleanup. |
For non-trivial runs, state Selected mode:, Why:, and Redirect: before
the first artifact or edit. For tiny direct work, one sentence can carry the
same information. Add a final Mode note: only when the user manually invoked
this skill, the request was ambiguous, or another mode/skill would fit better.
First Gates
Before reading more context, launching workers, or editing files:
- Treat the latest user message as authoritative.
Stop/pause/status-only/discuss-first language means read-only control mode
until a later message explicitly resumes execution.
- Check host goal state when available.
A
blocked or complete goal is a stop gate, not a resume cue. An active
goal is prior intent, not permission to override the latest user message.
- If this is an active-goal resume/debug turn, or a continuation of a
plan-backed durable run, use the shared Hot Resume and active capsule rules
in
../_shared/references/durable-run.md. Do not reread the full source
plan unless the capsule cannot answer a named next-action question.
- When changing
intuitive-flow itself, audit first and patch only the named
smallest delta after current-turn user permission.
For durable or resumed runs, read ../_shared/references/durable-run.md before
deciding whether to continue, relaunch, or stop. Read
references/context-budget-and-loop-guard.md only when the run needs the
Flow-specific experiment contract, self-modification guard, or blocker-loop
diagnostics.
Read First
Load only the reference needed for the selected route:
| Need | Read |
|---|
| Shared durable-run mechanics, Hot Resume, active capsule, control-plane/worker cadence, proof selector | ../_shared/references/durable-run.md |
| Flow-specific blocker experiment contract, self-modification guard, loop diagnostics | references/context-budget-and-loop-guard.md |
Source-of-truth, STATUS.md, CONTEXT.md, provenance, phase granularity | references/source-of-truth.md |
| Fuzzy idea shaping, single plan-file intake, unknown-unknown scout/reconciliation | references/plan-intake-and-autoplan.md |
| GSD ingest vs plan-phase routing, committed phase execution, changed-code cleanup scope | references/gsd-handoff.md |
| Whole-run preflight, goal ownership, soft continuation vs hard stop, checkpoint policy, tmux/goal/clear policy | references/checkpoints-and-auto-run.md |
Broad refactor route, semantic commits, final $intuitive-doc doc-alignment sub-phase, parked-todo closeout | references/refactor-and-closeout.md |
| Exact response and artifact templates | references/output-shapes.md and templates/ |
| Reference index | references/detailed-guidance.md |
If a route crosses multiple concerns, read the next reference just before that
boundary. Do not preload every reference.
Route Selection
- Routing precedence for overlapping prompts:
unknown owner or "what should we clean" ->
$intuitive-reduce-entropy;
known code/API/module seam -> $intuitive-refactor;
accepted direction without an execution contract -> $intuitive-preflight;
approved execution contract or tiny concrete task -> $intuitive-flow.
- Read-only diagnostics, status checks, and tiny concrete fixes: stay in the
main session with a short route note and focused evidence.
- Refactor, cleanup, stale API, compatibility, module layout, or known
architecture work: route to
$intuitive-refactor for accepted severities,
evidence ladder, and stop condition.
- Vague tasks, plan-backed work, broad durable implementation, or long-running
refactors: use durable Flow only after an approved
$intuitive-preflight
contract or equivalent approved execution contract exists. If the current
plan/conversation lacks context package, scope, non-goals, definition of done,
verification, route, worker strategy, and root-goal wording, route to
$intuitive-preflight instead of drafting a second Flow contract.
- Plan-backed implementation must have an approved preflight contract plus an
explicit unknown-unknown scout result or skip reason in the canonical plan,
unless the task is tiny direct work that is not using a plan as source of
truth.
- Canonical plan files use the fixed flat surface
docs/plans/<slug>.md.
Do not create lifecycle subdirectories under docs/plans/; record lifecycle
in the plan's ## Plan Ledger, keep docs/plans/README.md as the plan
dashboard when present, and use docs/status/active/<task-slug>.md for
compact active execution state. Maintain those docs by replacement and
pruning, not append-only history: stale next actions, superseded contracts,
and completed slice detail should be summarized or moved to linked evidence
notes before they distract from the current objective.
- Changed-code cleanup: use
$intuitive-refactor changed-code review on the
changed scope, then rerun relevant proof.
When the route is ambiguous, prefer the path that removes or reuses existing
entities over the path that adds a new one. If a proposal adds an entity, name
the existing surfaces considered and why they are insufficient.
For whole-flow or durable auto-runs, read
references/checkpoints-and-auto-run.md and identify the approved execution
contract. If none exists, route to $intuitive-preflight; do not maintain a
separate Flow-owned contract prompt.
Route Brief
Before non-trivial artifacts or edits, show a compact route brief. The full
shape lives in templates/route-brief.md; keep it short and route-specific.
Current state:
Latest user intent:
Host goal state:
Goal ownership:
Selected path:
Why:
Bypassed/left behind:
Execution surface:
Stop gate:
Stop/continue point:
For tiny direct work, one sentence is enough. Name plausible but skipped stages
such as $agent-planning-loop, grill-with-docs, unknown-unknown scouting
with gstack-autoplan, to-issues, GSD handoff, changed-code cleanup, or
verification.
Execution Surface
Keep the main session as route owner, integration point, and final verifier.
Use main-session edits for tiny direct work. For durable, stateful, long-running,
or parallel work, read ../_shared/references/durable-run.md,
references/checkpoints-and-auto-run.md, and the $skill-runner Codex
delegation reference before launching workers.
Hard Stops
Stop and ask the user when a choice would materially change product direction,
scope boundary, public contract, data model, roadmap ownership, security,
privacy, paid infrastructure, external-service dependency, destructive action,
broad file moves/deletes, unavailable local hardware/services, or locked
docs/ADRs. Also stop before expanding the accepted objective with a new durable
entity that was not in the approved scope.
If the next required evidence is blocked by a human action, local hardware, paid
account, external service, or another outside actor, stop instead of looking for
adjacent cleanup work.
Commit And Closeout
For significant implementation/refactor work, use
references/refactor-and-closeout.md and references/output-shapes.md for
semantic commits, final $intuitive-doc alignment, plan status updates, proof
cards, and parked-todo closeout.
Final responses for completed Flow implementation/refactor work must explicitly
include What changed, Proof, Scope changes, and Parked todos, even when
empty. Do not claim full completion when a required product-run, local/live, or
manual gate is skipped or blocked.