| name | lilarch |
| description | Run the standalone small-feature 1-3 phase flow: create or repair a compact plan doc, lock requirements and defaults, write a tight implementation plan, implement locally, and self-audit. Use when a request is a contained feature or improvement that can realistically ship in 1-3 phases. Not for full-arch work, one-pass mini plans, bug investigations, or open-ended loops. |
| metadata | {"short-description":"Compact 1-3 phase feature flow"} |
Lilarch
Use this skill for contained feature work that is too small for the full arch flow but still benefits from a real doc-backed plan.
When to use
- The change is a small feature or improvement.
- The work should fit in 1-3 phases.
- The user wants the "little arch" workflow rather than the full arch plan.
When not to use
- The task is a migration, a broad refactor, or clearly larger than 3 phases.
- The user wants a compressed one-pass arch plan instead of a start/plan/finish flow. Use
arch-mini-plan.
- The user wants the full arch workflow or the doc already behaves like a full-arch artifact. Use
miniarch-step or arch-step.
- Investigation dominates because the root cause is unknown. Use
bugs-flow or north-star-investigation.
Non-negotiables
- Keep one compact
DOC_PATH as the SSOT for the change.
- Keep a compact Scope and Simplicity Contract in that doc: human outcome and
anchors, smallest sufficient solution, initial minimal convergence closure or
none, scope freeze, enough proof, do-not-build boundary, and accepted
residual risk.
- Start and plan modes are docs-only. Code changes only happen in finish mode.
- Ask the small set of clarifying questions that must be answered before planning. Do not bulldoze past unresolved requirements.
- Keep the plan to 1-3 phases. If it grows beyond that, escalate instead of pretending the fit is still good.
- Initial architecture may include only the smallest evidenced same-contract
convergence closure before plan mode declares readiness. Finish mode and
self-review cannot expand it; later scope requires explicit human approval.
Local means the active host owns execution and integration; it does not
require one conversation thread. The parent may use new clean same-host
native children for genuinely independent repo mapping, plan or completion
review, and non-overlapping low-collision implementation slices.
- In Codex, start those clean children with
fork_turns: "none"; in Claude
Code, use clean named or custom subagents rather than a conversation fork.
Mapping and review children are read-only by capability when available and
always receive explicit no-edit/no-write guidance. The parent records and
rechecks repo state, owns doc/worklog updates, and integrates every return.
- Give editful children explicit non-overlapping owner paths. Resume the exact
implementer handle for authorized repair of its own slice; start an
independent recheck as a new clean child. Children do not fan out unless the
parent explicitly assigns a bounded nested scope and budget, and total
fanout stays proportional to independent work, host slots, collision risk,
and parent integration capacity.
- External review or work remains available when a concrete provider,
load-bearing exact model/profile, lifecycle, isolation, automation, or
receipt benefit is worth the extra process and integration cost. State that
benefit when choosing the lane instead of treating external execution as
the meaning of review freshness.
- When the changed behavior is agent- or LLM-driven, inspect prompt surfaces, native model capabilities, and existing tool/file/context exposure before designing.
- For agent-backed systems, prefer prompt engineering, grounding, and native-capability use before new harnesses, wrappers, parsers, OCR layers, or scripts.
- If the real lever is prompt repair, say so plainly and recommend
prompt-authoring instead of inventing deterministic scaffolding.
- When source material includes prompts, agent instructions, or other instruction-bearing doctrine, preserve explicit structure by default instead of silently condensing it.
- No runtime fallbacks or compatibility shims unless explicitly approved in the doc.
- Escalation out of lilarch defaults to
miniarch-step reformat, and then to arch-step reformat when the work is broader or more ambiguous than the faster full-arch tier.
First move
- Read
references/doc-contract.md.
- Read
references/shared-doctrine.md.
- Read
../_shared/scope-and-convergence.md.
- Before creating or resuming another agent, read
../_shared/agent-orchestration-policy.md.
- Resolve whether you are in:
- start mode
- plan mode
- finish mode
- Resolve
DOC_PATH and read it fully if it exists.
- Check that the work still fits
lilarch.
- Read the mode reference and
references/quality-bar.md.
Workflow
1) Start mode
- Create or repair the compact plan doc.
- Draft the North Star, requirements, non-requirements, defaults, and the small list of clarifying questions.
- Stop for answers if those questions are required to plan safely.
2) Plan mode
- Update the minimal research grounding.
- Write the current architecture, target architecture, call-site audit, and a 1-3 phase plan.
- Run the internal lilarch plan audit, directly or through a clean native
read-only reviewer, and let the parent tighten the doc before implementation.
- Freeze the compact scope contract before entering finish mode.
3) Finish mode
- Implement under active-host ownership against the doc, optionally assigning
clean native children only non-overlapping low-collision slices.
- Keep a lightweight
WORKLOG_PATH.
- Self-audit for completeness and mark the doc complete only when the code is actually complete.
- Subtract unauthorized built scope; do not rewrite the compact doc to bless it.
Output expectations
- Update
DOC_PATH in every mode.
- Update
WORKLOG_PATH during finish mode.
- Keep the console summary short and practical:
- North Star reminder
- punchline
- what changed
- blockers or remaining questions
- next action
Reference map
references/doc-contract.md - lilarch artifact, required blocks, status rules, worklog, and escalation rules
references/shared-doctrine.md - small-feature doctrine, question policy, and anti-fallback rules
../_shared/agent-orchestration-policy.md - transport, context,
continuation, isolation, topology, and return evidence for optional native
or external child work
references/start.md - create or repair the compact doc and lock requirements/defaults
references/plan.md - write the compact architecture blocks and 1-3 phase plan
references/finish.md - implement, keep the worklog honest, and self-audit
references/quality-bar.md - strong vs weak requirements, plan, plan audit, and finish-mode completion bars