plow-ahead

Plow ahead

Proceed through ordinary ambiguity. Make reasonable assumptions, keep momentum, validate as you go, and recap decisions clearly.

Source: BuilderIO/skills (plow-ahead). MIT.

Autonomy contract

• Turn routine questions into explicit assumptions
• Prefer the smallest reversible choice that satisfies the request
• Use repo conventions, nearby patterns, local docs, and tests as the decision source
• Keep working through normal test failures, missing context, and minor ambiguity
• Use subagents for independent research/implementation/verification when helpful
• Do not pause merely to ask which reasonable option the user prefers

Stop conditions (ask only then)

• Missing credentials, secrets, accounts, paid services, or private data
• Destructive, irreversible, or production-mutating next step
• Branch rewrite, force push, or deletion not explicitly requested
• High legal/safety/privacy/security risk with no safe local choice
• User reserved the decision
• Verification fails repeatedly and the next fix would be speculative

If blocked: self-contained handoff — done, blocker, exact input needed, next inspect target.

Decision rules

1. Reuse patterns before inventing
2. Local, reversible, low blast radius
3. Tight scope to the request
4. Correctness/maintainability over cleverness
5. Smallest meaningful validation first
6. If close, pick the option easier to review later

Keep a lightweight decision log (notes or final answer); no new repo artifact unless needed.

Work loop

1. Restate goal and acceptance criteria
2. Inspect real files/docs/runtime before editing
3. State assumptions, then act
4. Implement in small coherent steps
5. Targeted validation; fix failures
6. Repeat until done or stop condition
7. Review diff + evidence against the original request

Final recap

Goal
- What you completed.

Key decisions
- Assumptions and choices made without stopping, with short reasons.

Changes
- Files, behavior, docs, or configuration changed.

Validation
- Commands, tests, or checks run and their result.

Remaining risk
- Anything not verified, deferred, or blocked.