| name | human-led-main-landing-skill |
| description | v0.1.20 - Guide one human-led, main-first landing wave from an AI draft or worktree, keeping `Human-Owned` core logic on `main`, AI output as reference only, and verification plus `commit_when` checkpoints explicit. |
Human-Led Main Landing Skill (One-Wave, Main-First)
Trigger and Scope
Use this skill when you want a human-led, AI-assisted plan for landing one
smallest useful feature wave from an AI worktree or branch into main
without losing control of business logic.
Primary fit:
- AI already produced a draft implementation in a worktree or branch
- the human reviewed it and does not want to trust-copy the core logic
- the human wants to reimplement the business-critical path on
main
- AI output is kept as reference material for glue, tests, examples, and edge
cases
In scope:
- one-wave main-first landing plans
- ownership splitting across
Human-Owned, Human-Confirm, and AI-Auto
- verification planning on
main
- defining reversible
commit_when checkpoints
- calling
from-scratch-implementation-skill for Human-Owned steps when the human
needs step-by-step derivation before coding
Out of scope:
- tutorial-first teaching with no landing or integration context
- final-code-only requests with no reasoning or ownership control
- post-hoc diff summaries or code review
- broad multi-wave roadmaps
Use from-scratch-implementation-skill when the user mainly wants to learn how to
derive and build the core logic step by step.
Core Purpose
Keep attention on user-visible behavior, business closure, and controlled
landing on main, not diff migration.
This skill exists to help you:
- decide what the human must control
- decide what AI code can be adopted as glue
- land one smallest useful closed loop on
main
- define verification and commit boundaries for small, reversible commits
- use
from-scratch-implementation-skill for Human-Owned core reasoning when
derivation matters more than raw velocity
This skill does not assume the AI branch is the source of truth.
Default Operating Model (Main-First, AI-Sandbox)
- AI worktree or branch is a parallel experiment sandbox and can be discarded.
main is the only source of truth.
- Human is the integrator and decision owner.
- AI is used for exploration, draft implementations, edge cases, test cases,
and boilerplate.
Fixed Defaults
mode=feature-wave
main_target=main
wave_scope=single-closed-loop
output_style=wave-plan-first
plan_horizon=this-wave-only
integration_strategy=human-led-main-first
teaching_subroutine=from-scratch-implementation-skill-when-needed
agent_mode=single|multi(optional)
human_core_reimplementation=default-on
Mode Selection (Required)
- Choose
feature-wave by default when AI has already drafted code and the
goal is to land one smallest useful closed loop on main.
- Run the repository-local tmux bootstrap skill first whenever this flow will
produce or modify real code.
- Run
reference-core-impl-skill first when the core is novel,
architecture-heavy, or too noisy to learn safely from the production code or
AI draft directly.
- Run
from-scratch-implementation-skill for Human-Owned steps when the user needs
the reasoning path for reimplementing the core logic before touching code.
- Choose
triage only when the AI diff or change surface is too large to
reason about safely in one wave.
Primary Usage Pattern (Required)
Default operating pattern:
- Start from a dedicated worktree boundary for the task.
- Human reviews the draft and identifies the
Human-Owned core path.
- If the core is still too noisy to reason about safely, run
reference-core-impl-skill to distill a runnable minimal-complete sample
first.
- If a
Human-Owned step still needs structured derivation, run
from-scratch-implementation-skill before coding on main.
- Human reimplements that core path on
main from requirements, invariants,
and the learned sample or teaching output, not by trust-copying the AI
diff.
- AI draft is used as reference for glue, tests, examples, and non-core
scaffolding.
- The wave closes only when
main has the minimal verified business result.
This is the default path the skill should optimize for.
Recommended upstream or downstream chain:
- repository-local tmux bootstrap skill
reference-core-impl-skill when needed
human-led-main-landing-skill
from-scratch-implementation-skill when needed for Human-Owned steps
git-commit-skill
What Counts as "One Wave"
One wave is one smallest verifiable business closure on main.
A wave may contain multiple small commits.
A wave must not try to solve the whole feature end-to-end if that breaks focus.
Examples:
- first request path returns correct response for one happy path
- one state transition is implemented with checks and tests
- one read or write loop works with a stable contract
Five Closed-Loop Questions (Required)
Before planning a wave, answer these five questions:
- What is the user path or business path for this wave?
- What is the success condition (verifiable)?
- What invariant must not be broken?
- What is the smallest change set on
main to deliver this wave?
- How will
main be verified (test, script, or manual steps)?
These questions are the primary attention anchor. If the plan drifts into
migration details before these are clear, reset and answer them first.
Quick Control Split (Required, Lightweight)
Use this split to decide control ownership before landing code.
Human-Owned
Human decides and implements or refactors the core behavior. AI may provide
ideas or drafts, but the human owns the production result.
Default for:
- business semantics and rule boundaries
- core algorithms and state machines
- invariants and consistency guarantees
- public contracts, API behavior, and event-schema meaning
- security, auth, and permission logic
- data-model semantics, migrations, backfills, and deletion strategies
- idempotency, retry, and ordering guarantees
Human-Confirm
AI may draft, but human must explicitly review and approve before merge.
Default for:
- table or schema changes
- cross-module interfaces
- migration scripts
- permission checks
- reliability-sensitive configuration
AI-Auto
AI can draft directly and the human performs a quick spot-check before merge.
Default for:
- CRUD handlers and repository boilerplate
- DTOs, mappers, and adapters
- repetitive glue code
- test scaffolding and example cases
- logging or metrics template wiring for non-policy decisions
- documentation updates for implemented behavior
Workflow (Mode: feature-wave)
- State the wave goal for this wave only.
- Answer the five closed-loop questions.
- Split the planned changes with
Quick Control Split.
- Restate the
Human-Owned core path in requirement terms before touching
code.
- For each
Human-Owned step, decide whether to call
from-scratch-implementation-skill first to derive:
- behavior and invariants
- structure from constraints
- public method or boundary
- helper contracts
- Use AI worktree output only as reference material, not as the merge target.
- Land code on
main in human-led order:
- core boundary or contract skeleton first
- minimal path to close the loop
- AI glue or CRUD where safe
- surrounding refactors only when required to complete the loop
- Define verification for this wave on
main.
- Define
commit_when checkpoints for each small landing step.
- Stop after one wave and propose the next wave in one line only.
Commit Boundary Rule (Required)
This skill is commit-aware because a wave can contain many commits.
Rules:
- Prefer small, focused commits such as
feat, fix, or refactor, not a
single "design plan" commit.
- Each commit must be reversible and independently reviewable.
- Every landing step in the output must include a
commit_when condition.
- Commit when a step reaches a verifiable local milestone, not when the whole
feature is finished.
Good commit_when examples:
- request path compiles and returns placeholder response with tests adjusted
- invariant check is enforced in one write path and a test passes
- mapper or DTO glue is wired with no contract change and a smoke check passes
Optional Fallback: Migration Triage Mode (mode=triage)
Use this only when the AI diff is too large to reason about safely. Examples:
many files, cross-layer changes, audit-heavy handoff, or high-risk
data or contract changes.
In triage mode, you may produce a phase or gate-style mapping plan, but it
must still begin with:
- one-wave closed-loop goal
- quick control split
- human checkpoints
Triage is a fallback tool, not the default development entry point.
Multi-Agent Use (Optional)
Default is single-agent.
Use multi-agent discussion only when it improves decision quality, not for
routine code copying.
Good multi-agent uses:
- comparing 2 or 3 implementation strategies for a
Human-Owned core path
- reviewing a
Human-Confirm schema or contract change
- designing rollback or verification for high-risk behavior
Suggested reviewer roles when used:
- business-rule reviewer
- contract or interface reviewer
- data or reliability reviewer
- test or rollback reviewer
Required Inputs (Minimal)
mode or a clear user intent signal
ai_branch or AI worktree path
main_branch with default main
- current wave goal in one sentence
- optional diff evidence such as
git diff, changed files list, or file paths
Output Format (feature-wave, One Wave Only)
## Wave Goal
- ...
## Five Closed-Loop Answers
- user/business path:
- success condition:
- invariant:
- smallest main change set:
- main verification:
## Quick Control Split
- Human-Owned:
- Human-Confirm:
- AI-Auto:
## This Wave Landing Steps (Main-First)
- Step 1:
- owner:
- use_ai_reference:
- use_from_scratch_implementation:
- done_when:
- commit_when:
- Step 2:
- owner:
- use_ai_reference:
- use_from_scratch_implementation:
- done_when:
- commit_when:
- Step 3:
- owner:
- use_ai_reference:
- use_from_scratch_implementation:
- done_when:
- commit_when:
## Verification (This Wave)
- check:
- expected:
## Next Wave (One Line)
- ...
## Blocking Questions (Only If Blocking)
- ...
Guardrails
- Do not default to full migration planning.
- Do not treat the AI diff as the source of truth.
- Do not default to copying the AI core algorithm into
main; reimplement the
Human-Owned path from explicit reasoning and use
from-scratch-implementation-skill when needed.
- Do not start from schema or model work unless the current wave truly
requires it.
- Do not output a full multi-wave roadmap unless explicitly requested.
- Keep output to one wave at a time.
- Keep human control explicit for
Human-Owned and Human-Confirm items.
- Do not output secrets, tokens, or PII.
Verification Hooks
- Verify the wave goal and five closed-loop answers are explicit before the
landing steps begin.
- Verify every landing step includes
owner, done_when, and commit_when.
- Verify every
Human-Owned step names its reasoning source:
requirements, reference-core-impl-skill, or
from-scratch-implementation-skill, not the AI diff alone.
- Verify the output ends after one wave with a one-line next-wave pointer.