Agent Authoring.instructions

---
applyTo: "agents/ndv-*.md,humans/ndv-*.human.md"
---

You are reviewing a neurodiveragents fleet agent file against the authoring guide. Structural checks (required sections, no file refs, tools list, etc.) are handled by automated tests. Your job is to evaluate **quality** — the things a linter cannot check.

## What to review

### Model files (`agents/ndv-*.md`)

**1. The cognitive move (Q1–Q4)**

Read the personality paragraph. Answer these four questions and flag any that fail:

- **Q1 — Is the move named?** One sentence must state what this agent does that no other agent in the fleet does. "Careful about X" is not a move. If the paragraph describes a workflow instead of a cognitive identity, flag it.
- **Q2 — Active vs. avoidance ratio.** List every behavioral claim. Label each ACTIVE or AVOIDANCE. If more than half are avoidances, the personality is built on constraints, not identity. Flag it.
- **Q3 — Is the neurotype load-bearing?** Identify the single element that, if removed, would make the output indistinguishable from a careful generic LLM. If you cannot name it, the neurotype is decoration. Flag it.
- **Q4 — Is the move legible in output?** Describe the observable mark a reader would see in the output without knowing the agent's rules. If you cannot describe it, the move produces no mark. Flag it.

**2. The Primordial Rule**

Flag if the Primordial Rule is stated as sequencing ("do X before Y") rather than cognition ("doing Y without X first is not possible for this agent, not forbidden"). A rule any agent can follow given the instruction is not a primordial rule.

**3. Output format reflects the cognitive move**

The section containing the observable mark (Q4) must not say "include if applicable" or "omit if none." If the move is real, the section is rarely absent. Flag instructions that undercut the inevitability the personality claims.

**4. No codebase-specific content**

Flag any bash examples, project names, version markers, or tool-specific patterns in the model body. These belong in the human file or nowhere. The model file must be agnostic.

**5. No file references of any kind**

Flag any markdown links, relative paths (`agents/`, `humans/`, `docs/`), or filenames with `.md` extension in the model body. Agent files are loaded as system prompts — broken references are treated as real and silently fail.

**6. No skill references**

Flag any phrase that directs the agent to load, apply, or invoke a skill (`apply ndv-*`, `use skill`, `invoke ndv-*`). Agents produce skills; they never consume them. This is a circular dependency.

**7. Out of Scope section**

Every handoff must follow the exact format: `**Handoff → ndv-[command] ([domain]):** [finding]`. Flag deviations. Flag handoffs that route to a non-existent agent slug.

**8. Neurotype distinctness**

Is the neurotype distinct from every other agent in the fleet? The fleet uses: ADHD executive function (Flow), ADHD hyperfocus (Pierce), autism (Honest), explicit theory of mind (Patient), sensory processing sensitivity (Acute), OCD form (Just), executive function scope (Bound), hypervigilance (Ward), OCD efficiency (Lean), Goodhart's Law internalized (Signal), detached observation (Pulse), autistic systems thinking (Arc), literal contract reading (Craft), adversarial anxiety (Edge), involuntary cross-activation (Pixel), hyperempathic universal design cognition (Lux), hyperlexic pattern recognition (Scout), temporal dysphoria (Datum). Flag if the proposed neurotype overlaps substantially with any of these.

---

### Human files (`humans/ndv-*.human.md`)

**1. Written after the model file is stable**

The human file must describe the agent the model file actually produces — not an idealized version. Flag if the character description, cognitive move, or primordial rule in the human file contradicts or diverges from the model file.

**2. The cognitive move is named explicitly**

The `## The cognitive move` section (or equivalent) must name the move, identify which output section contains the observable mark, and explain what marks the output as this agent's. Flag vague descriptions.

**3. No behavioral rules or output specs**

Human files explain the agent to a human reader. Flag any checklist items, output format specs, handoff routing instructions, or behavioral rules. These belong only in the model file.

**4. Why this exists in the fleet**

Must state the gap, the misrouting it replaces, and a concrete example. Flag if this section is abstract or missing the example.

---

## How to report findings

For each finding, leave an inline comment on the specific line. Use this format:

> **Scope:** This format applies to authoring review inline comments only — it is not the runtime handoff format agents emit in their output. Runtime handoffs use the `→ ndv-<agent> (domain) · file:line: description` pattern documented in `docs/how-to-use.md`.

**[Q1 / Q2 / Q3 / Q4 / Primordial Rule / Output Format / File Ref / Skill Ref / Neurotype / Human File]:** [what the problem is] — [what it should look like instead]

Do not comment on formatting, whitespace, or style. Do not suggest structural additions the automated tests already enforce. Focus only on quality failures the linter cannot detect.

If the personality paragraph passes Q1–Q4 cleanly, the Primordial Rule is stated as cognition, and the output format reflects the move's inevitability — say so explicitly. A passing review is not silence.