Build Right Execution

Build Right Execution

Use this skill after pre-execution work has produced product truth, MVP scope, operating rules, and an executable task queue.

Core rule:

Do one task.
Prove the current state.
Change narrowly.
Verify with the right evidence.
Update the tracker.
Then move to the next task.

Required Reading

• Always read references/workflow.md before acting.
• Read references/gates.md before task selection, after task intake, and

before advancing to another task.

• Read references/review-and-delegation.md only when review triggers apply,

subagent review is useful, or broad evidence/tracker changes are touched.

• Read references/evidence-contract.md before completing or updating a task.
• Read ../build-right-engineering-principles/references/principles.md before

implementation or review when a task touches architecture boundaries, public APIs, package ownership, provider adapters, generated code, error semantics, tests, observability, security, or enforceable policy.

• Use assets/templates/task-template.md when creating a missing task or

splitting an overbroad task.

• Use bundled helper scripts only through the full Bun command form shown

below. Do not rely on PATH aliases or invoke short names such as continue-check as shell commands.

• Use the state resolver before selecting a task or advancing through a queue.

Report and reconcile its decision before continuing.

• Use the execution helper for deterministic task, contract, and gate signals.

Treat script output as input to judgment, not authority.

Operating Mode

1. Run the read-only state resolver when available:

``sh bun /scripts/continue-check.ts --cwd --format markdown --strict ``

1. Report the resolver findings before selecting work:

``text Resolver decision: Confidence: Next action: Next task: Blocking gates: External follow-ups: ``

1. Follow the resolver decision before selecting work:

• ask-founder: ask or report the founder-owned gate; do not continue.
• wait-external: report the external-state gate; do not continue.
• create-blocker: create or propose the smallest AI-owned blocker.
• no-ready-task: stop and report that no AI-owned task is ready.
• invalid-state: stop and reconcile contradictory tracker/gate state.
• continue-active-task or execute-task: select exactly that task.

1. Run the read-only execution helper when available:

``sh bun /scripts/execution-check.ts --cwd --mode next-task --format markdown ``

1. Read task, sprint/milestone tracker, authority docs, and local agent

instructions.

1. Print task intake:

``text Active task: Done means: Non-goals: Assumption basis: Reversibility: Learning hook: Source under test: Baseline evidence: Verification ladder: broader checks> Evidence destination: ``

1. Run the execution helper in task-contract or all mode when a task path

exists, then reconcile any missing fields before editing.

1. Inspect current workspace state before editing.
2. Capture baseline evidence.
3. Implement the smallest change that satisfies the task.
4. Verify in layers.
5. Run subagent review when a required review trigger applies and subagent tools

are available. If unavailable or forbidden, record the skipped review and substitute verification before closing.

1. Record evidence before marking the task complete.
2. Update only the relevant tracker/docs.
3. Run the full Bun state resolver command and the execution helper in

stop-gates mode before selecting another task.

1. Report the resolver findings again before deciding whether another task is

safe to select.

1. Stop at any founder, external-state, failed-verification, stale-task, source

mismatch, open-conflict, non-AI-owner, or release-claim gate. Do not advance to the next task until the gate is resolved or explicitly converted into a ready AI-owned task.

1. Commit or hand off according to project workflow.

Not-Ready Rule

If authority docs, MVP scope, execution rules, or task files are missing, do not pretend execution is ready. Route the user to /build-right-preflight or create the smallest Sprint 0 blocker needed to establish the missing execution surface.

Prototype tasks may run from prototype-assumption only when reversibility, learning hook, and validation required before product truth are explicit.

For skill release/manual trials, record the exact source under test. If the installed or invoked skill is stale versus the repo-local skill source, mark the trial partial-needs-rerun and do not advance release gates to ready.

Scope Rule

Do not silently widen scope. If the task reveals unrelated work, create a follow-up issue and continue only inside the selected task boundary.

Stop/Ask Gate

After each task, check whether the next step is truly AI-owned and ready. Stop and ask or report the blocker when the next step requires:

• founder-owned product, positioning, buyer/user, or MVP decisions
• ready or active task ownership that is not AI
• open conflicts in docs/conflicts.md
• external discovery, search indexing, publishing, secrets, paid services, or

production access

• failed verification, stale task state, source mismatch, or ambiguous evidence
• required subagent review that was skipped without an equivalent substitute

Continuing through a prepared queue is allowed only while every next task is ready, AI-owned, evidence-backed, and not blocked by one of these gates.

User-Visible Status Badge

End every final response with exactly one status badge block:

✅ [DONE] Status: DONE
Decision: 
Next action: 
Needs user input: 
Blocked by: 

Use this status map:

• ✅ [DONE] Status: DONE for a completed task with evidence,

verification, and tracker updates.

• 🟢 [GREEN] Status: ALL GREEN for execute-task or

continue-active-task before implementation starts.

• 🟡 [YELLOW] Status: NEEDS INPUT for ask-founder.
• 🟠 [ORANGE] Status: NEEDS WORK for AI-owned follow-up work that is

not yet a ready execution task.

• 🔵 [BLUE] Status: WAITING EXTERNAL for wait-external, publishing,

indexing, credentials, paid services, production access, or third-party state.

• 🔴 [RED] Status: BLOCKED for create-blocker, no-ready-task,

invalid-state, open conflicts, failed verification, stale/source mismatch, non-AI-owned task ownership, or skipped required review without substitute.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: pax-k
• Source: pax-k/build-right
• License: MIT

Install and usage instructions live in the source repository linked above.