# Grok 4.6 – A field guide (Eric Zakariasson)

Cursor / SpaceXAI engineer Eric Zakariasson’s launch-day X Article on using Grok 4.6 for a few weeks as a daily driver. Core thesis: communication + speed matter more than any single capability jump; highest leverage is a strict verification loop (“what done means”), not “work harder” phrasing.

Tags: source, grok-4-6, cursor, prompting, verification, agentic-coding, remotion, field-guide

## Who / what

| Field | Value |
|---|---|
| Author | Eric Zakariasson (@ericzakariasson) — Cursor / tinkering; ~80k followers |
| Form | X Article + announcement post |
| Title | Grok 4.6 – A field guide |
| Created | 2026-08-12T15:46:40Z (11:46 AM ET) |
| Engagement (~capture) | ~1323 likes · ~1173 bookmarks · ~118k impressions · 90 reposts · 65 replies |
| Post | https://x.com/ericzakariasson/status/2087566447178547494 |
| Article | https://x.com/i/article/2087564694706106372 |

Primary body: X API plain_text → raw/`grok-4-6-field-guide-2026-08-12/article.md` (cleaned structure; API joins noted).

## Bottom line (for Josh)

1. **Default model intuition:** Eric’s pitch is “good at all of it” + predictable shortcomings — not a spiky specialist.
2. **Prompting:** Drop “work very hard.” Prefer **short prompt + clear preference**, or a long spec when you know exactly what you want. Say **what done means**.
3. **Highest-leverage line:** force a verify→iterate loop until production-ready (browser use required for UI).
4. **Sync > async when you care:** speed+smarts pull him back to small synchronous steps instead of cold-reviewing huge async diffs.
5. **Steer where self-check is hard:** 3D, video, physics — give it a way to look (frames/screenshots) or you check.

## Key tips

### Information-dense communication

- Summaries carry real information (not task restatement).
- Short running updates are interruptible; quiet on small edits, narrates when touching many files.
- Still over-narrates sometimes (“we're working on”).

### Delightful speed → sync workflow

- 4.5 was fast; 4.6 is fast **and** smarter → more synchronous working style.
- Pattern: ask small → inspect → continue; same session can stretch into long-horizon work.
- Async still ships more while away, but loses the thread / cold big diffs. 4.6 pulls toward sync when the result matters.

### Short prompts, strict verification

- Phrasings like “work very hard” → **barely made no difference** (API wording; meaning: no/little effect).
- **Length** matters as specificity vs taste: long = you specify; short = model taste. With 4.6, short + preference often lands well.
- Long specs still fine (e.g. feedback widget).
- Spreadsheet A/B: two-page spec vs three sentences → nearly identical apps. The line that changed outcomes:

> Verify the function and design after implementation, and keep on iterating and verifying until it's production ready.

- That triggers open app → click real paths → check nested formulas → fix. Needs solid **browser use**.
- Hard-to-inspect outputs: “Improve the textures” failed; “capture the current frame, list what's wrong, fix only those” worked.
- Don’t tell it to keep pushing; define **done**. Comparisons used same prompt in isolated workspaces (4.5 vs 4.6).

### Where it needs steering

Almost all steering reduced to **ease of self-verification**:

| Domain | Why hard | Practical fix |
|---|---|---|
| Web UI | Easy — DOM + screenshot | Verification loop works well |
| 3D | Extra spatial dimension | Frame capture + list defects |
| Video | Time dimension | Sequence of frames + delta reasoning |
| Physics | Behavior over time | Give a way to look, or human checks |

### Why it’s his default

Wants a model he knows well across coding (interactive/visual + long repo sessions) and knowledge work (inbox, browser QA, click-through with no API). Not best-at-one-thing; good-at-all + known failure modes. Stay involved for motion/3D/polish; write acceptance criteria instead of trusting “done” summaries.

## Project experiments (4.5 vs 4.6)

| Project | Prompt / setup | Outcome |
|---|---|---|
| **AoE2-style** browser strategy | Economy, construction, combat, fog of war, objectives, readable HUD | 4.5: workable flat prototype. **4.6: isometric 3D + HUD/minimap first try** |
| **MSN Messenger** | Nostalgia recreation | Both know the reference; **4.6 more polish** (separate windows, winks) |
| **Excalidraw** presentation mode | Save/reorder named views; guided walkthrough; vague how-to; real OSS codebase | Same ballpark; **4.6 more first-pass detail** → fewer correction rounds. Skipping verify: summary said done but add-view broken; “run it and show me” found bad import |
| **Board deck** | Same fictional quarterly → both models | Gap is **presentation/hierarchy**, not analysis; 4.6 reads like a made deck |
| **Spreadsheet app** | Next.js + AI chat / Cursor SDK; 2-page spec vs 3 sentences | Nearly identical until verification sentence added |
| **Feedback widget** | Detailed spec | Long specs “still work fine”; repo linked: https://github.com/ericzakariasson/feedback-agent |
| **Cursor SDK Bridge** + **/rename-chat** | Launch posts; Remotion launch video; Bridge described with session capture, server handler, cloud agent dispatch | End-to-end sensible structure; may repeat components unless asked to break up *(API text join — see raw notes)* |
| **Remotion** film | 60–90s launch film for X TypeScript SDK from docs | Storyline + pacing; avoid all-at-once boxed uppercase titles. **4.6 more compelling**; video shows widest model spread |

Everyday work also cited: website navigation, **creating API keys via provider console clicks**, functional/visual QA, inbox triage.

## Relation to Grok 4.6 launch

Same-day (2026-08-12) official posts align with the field guide’s themes (long-running agents, interactive/visual strength, self-testing):

- Cursor: https://cursor.com/blog/grok-4-6
- SpaceXAI: https://x.ai/news/grok-4-6
- Positioning: builds on 4.5 for long-running agents + ambitious interactive/visual work; stronger first passes; more self-testing on long trajectories.
- Availability: Cursor, Grok Build, SpaceXAI API, OpenRouter, Vercel, Cloudflare.
- Promo: **2× included usage** in Cursor / Grok Build for first week.
- Pricing (official launch posts): **$2 / 1M input**, **$6 / 1M output**; fast variant ~2×.
- Secondary (VentureBeat / docs.x.ai — label carefully): 500k context; prompts ≥200k may bill at $4/$12; cached input noted at $0.50/1M on docs.
- Vendor eval headline: matches GPT-5.6 Sol on Artificial Analysis Intelligence Index (~61).

Entity page: [[grok-4-6]]. Distinct product: [[grok-bot]] (cloud teammates) — do not conflate model daily-driver tips with Bot product.

## Open questions

1. Exact preceding sentence for the Remotion / SDK Bridge API join (plain_text lost clause boundary).
2. Was the spreadsheet “both models” always 4.5 vs 4.6, or also other frontiers in some runs?
3. How portable is the verification one-liner outside Cursor browser-use (other harnesses)?
4. Fast-variant pricing/latency trade for Eric’s sync style — not discussed in the article.
5. Feedback-agent / SDK Bridge / rename-chat: public demos vs private; only feedback-agent GitHub linked in entities.

## Related

- [[grok-4-6]]
- [[grok-bot]]
- [[grok-bot-community-usage]]
- raw/`grok-4-6-field-guide-2026-08-12/`

## Sources

1. **Primary:** Eric Zakariasson, “Grok 4.6 – A field guide” (X Article) — API plain_text in raw/`…/article.md`
2. Announcement post metrics/meta — raw/`…/post-meta.json`
3. Cursor blog — https://cursor.com/blog/grok-4-6
4. SpaceXAI news — https://x.ai/news/grok-4-6
5. docs.x.ai model page — https://docs.x.ai/developers/grok-4-6 (secondary detail)
6. VentureBeat / DEV Community — secondary pricing/context only