
Implement in phases. The next step starts only after the previous one is verified.
SKILL // CHAIN STEP 04
The Rocinante corrects course at checkpoints instead of reaching its target on a single burn.A large change grows from checkpoint to checkpoint instead of arriving as one diff. /10x-implement executes an approved plan.md phase by phase: each one ends with verification, a manual confirmation gate and a commit, and the Progress section in the plan is the only state machine - an interrupted session resumes from the last green stage.
/10X-IMPLEMENT // EVERY PHASE IS A RETURN POINT
"Half the way under thrust, flip, half the way braking - and corrections along the way."
- a flight profile that can be aborted at any pointA change executed in one continuous burn has no places where it can be safely interrupted, verified or rolled back. /10x-implement cuts the execution into the plan's phases: a phase ends only when its success criteria are green, a human has confirmed the manual tests, and the change sits in a commit with its SHA written into Progress.
The philosophy: follow the plan's intent, adapting to the code you find. Plans are designed carefully, but reality gets messy - on divergence the skill stops and presents the problem (expected / found / why it matters) instead of quietly improvising.
for long or interrupted work, every phase is a return point: a commit plus green criteria instead of one giant diff at the end.

The agent executes an approved plan in small phases instead of one opaque batch of changes.
The input is plan.md from the change folder, optionally with a phase number (/10x-implement oauth-login phase 2). The skill reads the plan in full, every file the plan refers to, and lessons.md - the team's accepted rules shape every implementation decision in this run.
The output is a series of commits in the Conventional Commits convention - one per phase, in the form type(change-id): phase title (pN) - and a ticked-off Progress section in which every closed row carries the short SHA of the closing commit.

The plan is executed phase by phase, and the next step unlocks only after the previous one is verified.
on divergence from the plan, three exits, human decision: adapt and continue, skip this fragment, or stop and go back to planning.

A green verification closes the phase. Structural drift returns the decision to a person instead of triggering guesses.
The Progress section at the bottom of plan.md is the only source of truth about execution: no JSON on the side, no markers in comments, no helper file. State is not stored but derived: the first unchecked row = the next step; the current phase = the header above it; completion = the ratio of checked rows to all of them.
The skill mutates Progress only - the phase blocks (Overview, Changes Required, Success Criteria) are read-only. Mid-phase, checked rows without a SHA are a valid intermediate state; SHAs are added in bulk at the end of the phase, after the closing commit.
the answer always comes from parsing Progress - which is why an interrupted session resumes from the last green stage, not from zero, and the state cannot drift away from reality.

Every phase ends with verification and a separate saved state, so work can resume safely.
back to step 04 (implement) on the Core Skills Chain page - the context of the whole chain and the neighboring steps.

When a problem exceeds the approved plan, the agent stops and returns the decision to a person.
NO RESULTS // TRY ANOTHER QUERY