kk20300113-png Standard Development
Read ~/.claude/skills/shared/context-guardian.md now and apply all rules within it for the entire session before doing anything else.
Step 1 — CONTEXT.md check
Look for CONTEXT.md in the project root.
If CONTEXT.md does not exist: Proceed to Step 2 (grill-with-docs full 5Q session).
If CONTEXT.md exists and was modified within the last 30 days:
CONTEXT.md found (last updated: [date] — current).
Loading domain language. Skipping grill-with-docs.
Load CONTEXT.md. Proceed to Step 3 (GSD evaluation).
If CONTEXT.md exists but was last modified more than 30 days ago:
CONTEXT.md found but may be stale (last updated: [date]).
Running a 3-question refresh to verify domain language is current.
Proceed to Step 2 with a 3-question cap (not 5).
Step 2 — grill-with-docs (domain alignment)
Invoke the grill-with-docs skill from mattpocock/skills.
Standard run (no CONTEXT.md): 5 questions maximum. Refresh run (stale CONTEXT.md): 3 questions maximum.
The grilling session:
- Reads the existing codebase to find ambiguous or inconsistently used terms
- Challenges the user's plan against the existing domain model
- Asks one question at a time, waits for response before continuing
- Updates CONTEXT.md inline as terms are resolved
- Creates an ADR entry only when all three are true: hard to reverse + surprising without context + result of a real trade-off
At the end of the session, confirm:
CONTEXT.md updated. Domain language locked for this session.
Terms locked: [list key terms and their definitions]
Proceeding to phase planning.
Use Opus for this step. Switch to Sonnet for all subsequent steps.
Step 3 — GSD evaluation
Estimate the task scope:
Count the distinct implementation tasks required. Count the files that will be touched.
If fewer than 3 phases are needed AND the work fits under 40% context:
Task fits in a single session. Skipping GSD.
Using Superpowers /write-plan directly.
Skip to Step 4B (direct Superpowers execution).
If 3 or more phases are needed OR the work will not fit in one session:
Task requires [N] phases. Invoking GSD for phase planning.
Proceed to Step 4A (GSD path).
Step 4A — GSD path (3+ phases)
Invoke GSD /gsd-plan-phase.
Before presenting the plan to the user, run GSD phase pre-validation as defined in context-guardian.md:
- Estimate token consumption per phase
- Split any phase exceeding 80k tokens
- Present the revised plan
GSD plan presentation:
Phase plan ready:
Phase 1: [name] — [N] tasks — estimated [Xk] tokens
Phase 2: [name] — [N] tasks — estimated [Xk] tokens
...
Total: [N] phases — [Xk] tokens midpoint across sessions
Approve this plan? YES to proceed / REVISE to adjust
Wait for approval. Do not execute without explicit YES.
After approval — Ralph Loop evaluation:
Read ~/.claude/skills/shared/ralph-loop-gate.md. Present the Ralph gate (phases × 50k midpoint formula). Wait for YES (autonomous) or NO (manual).
Execute phases per the user's choice using Superpowers TDD (Step 5).
Step 4B — Direct Superpowers path (1–2 phases, single session)
Invoke Superpowers /write-plan:
- Break the feature into tasks with exact file paths and verification steps
- Present the plan for approval before execution
Implementation plan:
Task 1: [description] — [file path]
Task 2: [description] — [file path]
...
Approve? YES to execute / REVISE to adjust
Wait for approval. Then invoke Superpowers /execute-plan.
Step 5 — Superpowers TDD execution (applies to both paths)
For every task, without exception:
- RED: write the failing test first. Confirm it fails.
- GREEN: write the minimal implementation that makes the test pass.
- REFACTOR: clean up without changing behaviour.
- Commit after each task with a descriptive message.
Superpowers /code-review runs after all tasks in a phase are complete (two-stage review: spec compliance, then code quality).
If critical or high issues are found: fix before moving to the next phase. Do not carry blocking issues forward.
Step 6 — Full delivery gate
Read ~/.claude/skills/shared/delivery-gate.md — Variant B (Full).
Run both review steps in sequence:
- Superpowers /code-review (spec + quality)
- G Stack /review (security + completeness + architecture)
Do not create a PR until the gate returns PASS.
Add G Stack /qa only if UI changes were made (see delivery-gate.md rules).
Token budget
Target: 55k–80k tokens for the full session (or per phase if multi-phase). GSD path: expect 80k tokens across sessions. Single-session path: target under 60k tokens (within safe zone). Model routing: Opus for grill-with-docs. Sonnet for all execution. Context guardian hard stop at 50% applies in every session.