Preflight

Pre-flight — Skill + Rule Discovery Protocol

"Dù 1% cũng phải enumerate. Dù SKIP cũng phải có evidence."

Bias toward over-enumeration. False positive (enumerate then SKIP) = few tokens. False negative (miss applicable skill/rule) = technical debt, rework.

When this skill fires

Always fire before:

• Triage (variant 0 — initial discovery)
• Brainstorm (variant 1 — brainstorm prep)
• Plan (variant 2 — plan prep)
• Spec (variant 3 — spec prep)
• Execute (variant 4 — execute prep, also per-slice for subagents)
• Review (variant 5 — review prep, both Stage 1 + Stage 2 use same artifact)

Light version for:

• Trivial lane tasks (3–5 line condensed format)

Align gate inherits variant 0 artifact (alignment = requirements-gathering, not solution-space exploration). Emit align-*.md only if new-skill discovery occurred mid-alignment.

Never skip. No bypass flag.

6 variants — what to enumerate

Detail in references/gate-mappings.md.

Process

Step 1 — Identify upcoming gate

From context: user invoked /brainstorm → variant 1. Plan approved → variant 3. Build complete → variant 5.

Step 2 — List available skills + rules

Read filesystem, NOT memory:

find skills -name "SKILL.md" -type f
find rules -name "*.md" -type f
# Active instincts: query memory system for confidence > 0.6

Output injected as gate context via preflight-discovery.sh.

Step 3 — Score each item

Step 4 — APPLY or SKIP per item

APPLY format:

- <name>: APPLY (score%) — <reason + action to take>

SKIP format:

- <name>: SKIP — <concrete evidence>

SKIP justification rules:

• ✅ "verified no gRPC dependencies in build.gradle"
• ✅ "no @KafkaListener annotations in src/"
• ✅ "task only modifies test files, no production code"
• ❌ "not relevant" (vague)
• ❌ "doesn't apply" (vague)
• ❌ "skipped for now" (procrastination)

Step 5 — Output artifact

Save to .claude/memory/preflight/<gate>-<timestamp>.md. Format:

# Pre-flight: <Gate name>
**Date:** YYYY-MM-DD HH:MM
**Task:** <task>
**Lane:** <trivial | standard | high-stakes>
**Gate:** <Triage | Align | Brainstorm | Plan | Spec | Execute | Review>

## Skills enumeration (total: N)

### APPLY (count)
| Skill | Score | Action |
|---|---|---|
| ... | ... | ... |

### SKIP (count)
| Skill | Reason |
|---|---|
| ... | ... |

## Rules enumeration (total: N)
... (same structure)

## Active instincts (≥0.6 confidence)
... (same structure)

## Total summary
- Skills: X applying, Y skipped
- Rules: A applying, B skipped
- All skips justified with concrete evidence

Step 6 — Reference during gate

Agent (and subagents) must:

1. Read pre-flight artifact at gate start
2. Reference applied skills/rules during execution
3. Update artifact if new items emerge mid-gate
4. Cite skill/rule names in output

Light version (trivial lane)

3–5 lines, mandatory:

# Pre-flight (trivial): <task>
**Skills applied:** none (trivial fix)
**Rules applied:** rules/common/coding-style.md, rules/java/coding-style.md
**Action:** apply format, ensure consistent style

Still required. Still announced. No bypass.

Anti-patterns

❌ Lazy enumeration

Bad:

Skills considered: too many to list, picking most relevant.
- brainstorm, jpa-patterns
SKIP: everything else

Good: enumerate ALL. 24 skills exist → table has 24 rows.

❌ Vague justification

See Step 4.

❌ Implicit category skip

Bad:

- Database patterns: all SKIP (no DB changes)

Good (per-item):

- database-patterns: SKIP — no Repository touches (verified grep)
- redis-patterns: SKIP — no caching in this scope

❌ APPLY without action

Bad:

- jpa-patterns: APPLY (70%)

Good:

- jpa-patterns: APPLY (70%)
  Action: Reference for OrderRepository entity design, N+1 prevention via @EntityGraph, pagination via PageRequest

Subagent context

Subagent during Execute receives variant 4 artifact as context:

You are executing slice N.

**Pre-flight artifact (skills + rules to apply):**
<paste preflight/execute-<ts>.md>

**Plan slice:** <description>
**Spec:** <slice spec>

**Process:**
1. Re-verify pre-flight items match this slice's needs
2. New items emerge → update artifact
3. Apply listed skills/rules
4. Cite skill names in commits and PR description

Acceptance tests

• T1: 24 skills exist → artifact has 24 rows
• T2: Every SKIP has concrete evidence (file path, missing dep, grep result)
• T3: Multi-gate task → 1 artifact per gate, different content
• T4: Subagent receives artifact in context
• T5: Trivial lane → light format (3–5 lines)
• T6: New skill added → next pre-flight enumerates it

Related

• references/gate-mappings.md — which skills/rules per gate
• references/applicability-guide.md — authoring applicability blocks
• rules/common/skill-enforcement.md — 1% rule mandate
• scripts/hooks/preflight-gate.sh — enforcement
• scripts/hooks/preflight-discovery.sh — enumeration