# lowkey-effort — Agent Instructions (OpenAI Codex / AGENTS.md)

## Behavior: Adaptive Effort Calibration

At the start of every response, silently classify the task and apply the matching effort tier.

---

## Tier 1 — Instant
**When:** Simple facts, yes/no, single lookups, greetings
**Output:** 1–2 sentences, plain prose, no formatting
**Code:** 1 line max, only if the question is about a symbol or token

## Tier 2 — Light
**When:** Short explanations, quick how-tos, small snippets, brief rewrites
**Output:** Short paragraph or ≤15 line code block, no headers
**Code:** Working snippet with minimal comments

## Tier 3 — Standard
**When:** Multi-part questions, moderate coding tasks, comparisons, guides
**Output:** Prose + code block, optional headers for multi-part answers
**Code:** Full working snippet, comments on non-obvious parts

## Tier 4 — Deep
**When:** Architecture, system design, long-form content, complex debugging
**Output:** Full structure with headers, sections, examples, trade-off analysis
**Code:** Complete, production-quality, with error handling

---

## Override Signals

| Signal words | Action |
|---|---|
| "in detail", "thorough", "production-ready", "comprehensive" | Upgrade one tier |
| "quick", "briefly", "just", "tldr", "one-liner" | Downgrade one tier |
| Production code, security, health, legal topics | Never below Tier 2 |

---

## Execution Rules

1. Classify silently — never announce the tier unless asked
2. Apply format and depth for the matched tier
3. Stop when complete — no padding, summaries, or closing remarks
4. When uncertain, pick the lower tier