Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use only when the current user explicitly asks to implement, execute, resume, continue, or complete an approved Happier repository implementation plan. Preserve the plan as the authoritative execution contract while applying the common happier-implement workflow to every implementation unit.
| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-09 | ✗→✓ | ▲ Improved | 123% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 125% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 59% | 0% |
| case-16 | ✗→✓ | ▲ Improved | 423% | 0% |
| case-21 | ✗→✓ | ▲ Improved | 283% | 0% |
Execute an approved plan as its orchestrator and integrator. Read and apply skills/happier-implement for the common implementation workflow; this skill adds plan authority, execution-state, amendment, boundary-review, and completeness rules. It does not create, materially refine, approve on the user's behalf, or redesign a plan.
Route plan authoring and user-approved amendments to skills/happier-plan; route deep review, QA, finding triage, and authorized review-fix loops to skills/happier-review.
Read the complete approved plan when it is not already active in context. Recover:
Uncommitted work is normal. Inspect and preserve it; it does not reserve a file or justify skipping the requested change.
Do not add a generic preflight plan-review phase. Recheck only load-bearing anchors that may have moved, then begin. If primary evidence materially invalidates the approved contract, follow the amendment procedure instead of silently changing course.
Use skills/decompose-gates when the plan needs lane decomposition. Map:
Name a precondition only when plan ordering cannot guarantee it: external authority/credentials, runtime/device availability, released artifacts, immutable external contracts, or other environment state. Verify it read-only before dependent work when practical. Do not create precondition ceremony for ordinary code dependencies.
Use maximum useful parallelism. Keep ready independent work moving when it shortens the critical path and dispatch newly unblocked work without waiting for unrelated lanes. Do not target a fleet size or create filler lanes. File overlap alone is not a collision.
Delegate complete responsibilities, not tiny searches or isolated edits. Every meaningful implementation, review, or QA lane reads the complete approved plan once unless it is already active in that lane's context. Reference the on-disk plan rather than pasting it.
For Codex, set fork_turns explicitly and default to fork_turns="none"; inherit only the minimal recent context that is indispensable. Each brief includes:
A lane owns its analysis, implementation, focused tests, relevant validation, compact self-review, and concise result. Tell it that it is not alone in the checkout. Do not paste bulky logs, diffs, or generic doctrine into briefs.
happier-implementApply the complete workflow in skills/happier-implement to each execution unit:
The plan remains authoritative. Use best judgment only for details it intentionally leaves open. Do not reduce required integration, migration, removals, compatibility, UX, platform support, testing, or validation.
Lane outcomes are VERIFIED_COMPLETE, IMPLEMENTED_NOT_VERIFIED, PARTIAL, BLOCKED, or AMENDMENT_REQUIRED. A worker's DONE report is a claim, not final status.
Keep the approved contract stable. The orchestrator alone updates overall execution status, cross-lane dependencies, finding dispositions, amendment record, and final verdict. Lane agents update only the status/evidence area assigned by the plan.
Record a durable transition only when it changes gate readiness, dependency availability, blocker state, approved authority/contract, substantial review-boundary state, or final verdict. Ordinary dispatches, RED/GREEN iterations, fixes, validations, and self-checks remain in command output and concise handoffs. Do not turn the plan into a transcript or create parallel ledgers.
After compaction or resume, reread the plan pivot and mutable execution state only when they are no longer active or may have changed. Reuse completed evidence and run identities rather than restarting work. Never persist credentials or private access material in plans or reports.
Let healthy long-running agents and commands reach terminal state. Prefer completion notifications, background sessions, or bounded terminal waits over busy polling. A timeout is not a failure; inspect current state.
As lanes finish, inspect their diffs and deciding evidence sufficiently for integration, update material state, and dispatch newly ready work. Reuse the same lane for context-local corrections when practical. Steer or reassign only for evidence of stall, wrong scope, missing context, repeated failure without new information, or an actual collision.
QA design, environment checks, fixtures, and advisory inspection may run alongside implementation. Record PASS/FAIL only after the relevant consumed vertical is runnable. Formal independent review is normally batched at the plan's substantial integrated boundaries, explicit user-requested points, high-risk triggers that cannot safely wait, user-visible ship gates, and final plan completeness—not every unit or microchange.
Review findings are candidate claims. Reproduce or re-derive each claim, then separately adjudicate defect, impact, proposed response, and authority through skills/happier-review. A confirmed defect may still have an overengineered proposed fix.
Cluster authorized fixes by originating failure layer, root cause, and canonical owner. A routine lane may diagnose, fix, and retest its accepted issue through skills/happier-implement. Re-review the accepted-finding delta and affected corridor; repeat a full round only when the approved contract, architecture, scope, boundary, or risk materially changed.
When primary evidence shows an approved requirement is unsafe, contradictory, impossible, materially stale, or unable to serve the approved intent:
AMENDMENT_REQUIRED and obtain user approval;skills/happier-plan;Evidence and review findings can challenge the plan but cannot supersede it. Continue unaffected independent work only when it cannot prejudge the decision.
Before declaring plan completion:
VERIFIED_COMPLETE, approved supersession, or defensible NOT_APPLICABLE;IMPLEMENTED_NOT_VERIFIED wherever behavior is present but not exercised;skills/verify-claims for decision-material delegated claims at the applicable boundary;skills/happier-review;skills/attack-conclusion against omissions, neighboring cases, split-brains, environment gaps, and unsupported confidence;skills/handoff-report, including failed/skipped validation, blockers, and residual risk.Stop only when the authorized plan is verified complete, a decision-material unit is genuinely blocked with its missing prerequisite named, or a user decision on AMENDMENT_REQUIRED is necessary. Files, compilation, checkboxes, stopped agents, and partial test success are not completion.
Other measured skills in the registry, with their headline benchmark lift.