Instruction Creator

Author instructions that pass instruction-review's five gates by construction (plus Gate 6 for always-on instructions). The gates are the quality bar; invert them, and you build to standard the first time.

Mirrors skill-creator's seven-phase structure with instruction-specific guidance. If a phase here disagrees with skill-creator on the shared scaffold, skill-creator wins; this file owns only the per-type guidance.

When to Use

• Authoring a new instruction in .github/instructions/
• Refactoring an existing instruction (scope shift, applyTo recalibration, split, merge)
• Promoting a Mall instruction into your heir's brain

When not to use

• Authoring a skill (a procedure with steps) → use skill-creator
• Authoring a prompt (a slash-command workflow) → use prompt-creator
• Authoring an agent (a delegated sub-process role) → use agent-creator
• A one-off rule for a single file with no recurring trigger → not a brain artifact

The Seven Phases

Each phase inverts one of the five (or six) gates. Author against the phase, pass the gate.

Phase 1 — Intent capture

Answer in writing before any drafting:

1. What rule does this enforce? One sentence. If it contains "and" / "+", you have two instructions — split.
2. What is the trigger condition? Concrete — drives the applyTo glob in Phase 3. "When working on Azure code", "On every turn", "On files in AI-Memory/".
3. What does the agent do differently when this fires? Be specific. If the answer is "be more careful", the rule is too vague.

If the rule is a procedure (multi-step operation) rather than a condition-action pair, stop — author a skill, not an instruction.

Phase 2 — Prior-art scan

Select-String -Path .github/instructions/*.instructions.md -Pattern "<keyword>"

Phase 3 — Draft against Gate 1 (Spec) and calibrate applyTo

Frontmatter template:

---
description: "<third-person, what + when, ≤1024 chars>"
applyTo: "<glob pattern>"
lastReviewed: YYYY-MM-DD
---

applyTo calibration is the load-bearing decision for instructions. Too broad = token waste on every turn; too narrow = misses the cases it should catch.

If applyTo: "**", the body must include an explicit always-on rationale section (Gate 6 requirement).

File location: .github/instructions/<kebab-name>.instructions.md (flat — no subfolders).

Phase 4 — Draft against Gate 2 (Quality)

Phase 5 — Draft against Gate 3 (Scope)

Phase 6 — Draft against Gate 4 (Safety)

Same checks as skill-creator Phase 6. Instructions rarely bundle scripts, but if they reference scripts, the script-discipline rules from skill-creator apply.

Phase 7 — Dogfood self-audit (with Gate 6 if always-on)

Before committing, run instruction-review's five gates (plus Gate 6 for always-on). Verdict lives in the commit message.

Gate 6 check — only for applyTo: "**":

• Body ≤150 lines
• Explicit "always-on rationale" prose names why this fires every turn
• No duplication of content owned by a skill

Anti-Patterns

Falsifiability

This skill's design has failed if any of the following occur within 90 days:

• ≥2 instructions authored using this guide fail instruction-review Gate 1 or Gate 6 on first self-audit (phases didn't internalize the gates)
• No new instructions authored via this guide in 90 days (decorative, not load-bearing — sunset)
• applyTo calibration table produces consistent miscalibration (always-on chosen where pattern would do, or vice versa) ≥3 times in a quarter

Track as you would any falsified discipline (commit log, retraining notes, or curation ledger if your repo ships one) tagged [INSTRUCTION-CREATOR-MISS].

Related

• instruction-review — the five gates (+ Gate 6) this skill inverts
• skill-creator — sibling for skills; canonical seven-phase scaffold
• prompt-creator — sibling for prompts
• agent-creator — sibling for agents
• act-pass — required for medium-stakes instruction authoring