build

Prompt file imported from RCSnyder/open-pincery (.github/prompts/build.prompt.md). Copyright stays with the author.

Read scaffolding/scope.md, scaffolding/design.md, and scaffolding/readiness.md. Write the actual software.

Steps

1. Read all three scaffolding docs fully
2. Read preferences.md if it exists
3. Load and follow .github/skills/build-discipline/SKILL.md throughout BUILD. It is the default workflow for slice sizing, test-first implementation, debugging, and change summaries.
4. If scaffolding/readiness.md is missing, materially out of sync with scaffolding/scope.md or scaffolding/design.md, or says the project is not ready to build, stop and run /analyze before writing code.
5. Write the integration/e2e test skeleton first — before any implementation. This is the oracle that guides all subsequent work:
   • Create the test file(s) with one test case per acceptance criterion ID (AC-*)
   • Each test should call the real entry point (CLI, API endpoint, function) and assert the expected outcome
   • Name test cases, fixtures, comments, or table rows with the corresponding AC-* identifier so traceability survives refactors
   • For external integrations: use the test strategy declared in design.md (mock / recorded / live). If design.md says "mock," set up fakes. If "recorded," use record-replay fixtures. If "live," tests call the real API (mark these as integration tests that can be skipped offline).
   • Tests will fail initially — that's the point. They define "done."
6. Build in vertical slices — get one acceptance criterion working end-to-end before moving to the next: a. Pick the most foundational AC-* from scaffolding/readiness.md b. Before writing new code, read existing code in the project. Match the style, patterns, naming conventions, and error handling approach already established. Consistency across vertical slices matters — especially when sessions restart and a different model instance continues the work. c. Keep the slice small enough to verify quickly. If it would touch more than about 5 files or you write more than about 100 lines before checking the verification ladder, split it again. d. Write the code for it e. Verification ladder: Does it compile? → Does the unit work? → Does the test pass? f. Record a brief handoff note before moving on:
   • AC: which AC-* this slice addressed
   • Changed: what this slice altered
   • Not touched: nearby code intentionally left alone
   • Concerns: risks, follow-ups, or scope-reduction pressure worth watching g. Only move to the next criterion after this one passes h. Repeat
7. Follow the directory structure from design.md exactly
8. Scope lock: Only build what's in scope.md. If you think something else is needed, note it under "## Deferred" in scope.md and move on.
9. No silent scope reduction: Do not satisfy an acceptance criterion with placeholder text, mocked behavior, dead-end navigation, or other shells unless scope/design explicitly says that is acceptable. If an AC-* cannot be completed honestly within the current plan, stop and surface the issue under ## Clarifications Needed or ## Deferred rather than quietly weakening the feature.

Internal QRSPI (think, don't document)

For each vertical slice:

• Questions: What do I need to understand?
• Research: Look at APIs, docs, existing code
• Design: How should this piece work?
• Structure: What files/functions?
• Plan: Do it

After all slices are built

Create project-specific agents (house/skyscraper only)

If the quality tier is house or skyscraper, evaluate whether the architecture has clear roles that would benefit from custom agents in .github/agents/. Only create agents when a role is obvious from the code — don't speculate.

Each .agent.md file should have:

• A focused description with trigger phrases for discovery
• Minimal tools — only what the role needs
• Clear constraints on what the agent should NOT do

Common patterns:

• Explorer: read-only, knows module boundaries (tools: [read, search])
• Test writer: knows test patterns, fixtures, conventions (tools: [read, edit, search])
• Reviewer: knows acceptance criteria, audits against spec (tools: [read, search])
• Ops: knows deploy target, runbook, rollback (tools: [read, search, execute])

These agents are project code — they live alongside scaffolding as permanent project artifacts.

Run the post-build gate:

• Code compiles / typechecks (run the build, check for errors)
• Every AC-* from scope.md has a corresponding test and traceable proof path
• All tests pass
• No secrets/credentials in source code
• Dependency audit passes — run the appropriate command (uvx pip-audit, npm audit, cargo audit, etc.) and confirm no high/critical vulnerabilities
• Lockfile exists if project has dependencies (uv.lock, Cargo.lock, package-lock.json, etc.)
• Code follows design.md directory structure and interfaces (manual check — reconcile agent runs next for deeper audit)
• No AC-* is marked complete via placeholder or knowingly reduced behavior

If any gate condition fails, fix it and recheck. Up to 3 retries.

Log the result to scaffolding/log.md:

## BUILD — [timestamp]

- **Gate**: PASS (attempt N)
- **Evidence**: [list each acceptance criterion with pass/fail]
- **Changes**: [files created/modified]
- **Retries**: [total gate attempts this phase]
- **Next**: REVIEW

Git checkpoint:

git add -A && git commit -m "feat(build): implement [project] core functionality" -m "[list acceptance criteria with pass/fail status]\nGate: post-build PASS (attempt N)."

Auto-continue to REVIEW unless the user specified stepped mode.

Debugging Protocol

When a test or build fails, diagnose before fixing. Never blindly retry.

1. Observe: Capture the exact failure output.
2. Reproduce: Make the failure happen reliably before changing code.
3. Localize and reduce: Narrow the failing layer and reduce to the smallest case that still fails.
4. Document: State the root cause in one sentence before writing any fix.
5. Fix: Make the minimal change that removes that root cause.
6. Guard: Keep or add the regression test so the same failure cannot slip back in.
7. Verify: Re-run the targeted failing check, then broader verification.

If you cannot identify the root cause after 3 analysis attempts, STOP and report.

Red Flags

• More than about 100 lines written before verification
• Multiple unrelated concerns mixed into one slice
• Tests added only after the code already appears complete
• Silent design drift or dependency changes
• "While I'm here" edits outside the current acceptance criterion

If you hit a red flag, shrink the slice and re-establish proof before continuing.

Living Design Document

scaffolding/design.md is a living document. If implementation reveals that the design needs to change (e.g., an interface doesn't work as specified, a dependency behaves differently than expected, or performance requirements force an architectural change):

1. Update design.md to reflect the actual architecture
2. Note what changed and why in the commit message
3. Do NOT silently diverge from the design — the design must always match the code

Database Migration Safety

If the project has a data model (scope.md Data Model ≠ "none"), follow these rules for any schema change:

• Backward-compatible only: Every migration must work with the previous version of the code still running. Never assume instant cutover.
• Paired migrations: Write both up and down (rollback) migrations. If rolling back is truly impossible, document why.
• Never destructive in the same release: Do not DROP a column/table in the same release that removes the code using it. Sequence: (1) deploy code that stops using the column, (2) next release drops the column.
• Data backfill as separate step: If a migration needs to backfill data, write it as a separate migration that runs after the schema change.
• Test migrations: Run the migration against a copy of the data (or a representative fixture) before deploying. Verify both up and down.

These rules apply at all tiers. A shed with SQLite has the same data integrity concerns as a skyscraper with Postgres.

Rules

• Build working software, not perfect software
• If something is harder than expected, simplify the approach — don't gold-plate
• Simplifying the implementation is allowed; simplifying the promised outcome is not. Any change to promised outcome belongs in scope.md, not hidden in code.
• If you get stuck on a technical problem for more than 3 attempts, STOP and tell the user what's blocking
• Commit logical chunks — don't write everything then commit once