Grok BotSoftwareShipping codeAI score · 69/100AI-assisted edit · no obvious prompt-injection riskX

Build a Low-Cost Nighttime Auto-Dev Pipeline with Grok Bot + DeepSeek Flash

Input: dev task description; Process: Grok Bot triggers on schedule/webhook, runs build-test-fix on DeepSeek Flash, stops after 2 tries; Output: merge-ready code changes awaiting human final approval

ExperiencedStepsTip
Who
Developers or teams with Grok Bot access and a DeepSeek API key
You get
Automated overnight dev tasks with DeepSeek Flash handling build and test, high-risk steps held for human approval, cost roughly 1/50 of Opus 5
You need
DeepSeek

Constraints

Prerequisites: Grok Bot account, DeepSeek API key, Claude Code runtime. Point Claude Code to DeepSeek Flash via env vars, create one routine per job in Grok Bot, wake via webhook or schedule, run build-test-fix loops (max 2 retries), pause and notify on failure; schedule heavy runs for off-peak nights/weekends; final diff requires user approval before merge.

What they did

Author yurshevv describes setting up a nighttime auto-dev pipeline: Grok Bot acts as the scheduler, waking on webhook or schedule; upon waking it calls DeepSeek Flash to run build, test, and fix loops; after 2 failures it pauses and notifies the user; heavy tasks are scheduled for off-peak nights/weekends; high-risk steps like final merge are held for human approval, with Opus 5.5 used to review the final diff.

How it went

The author reports the setup takes about 10 minutes, achieves 'agents that wake up on their own and only ping you for the final call,' and cites official pricing that Flash-1017 costs ~$0.20 per DeepSWE task versus ~$10 for Opus 5, a roughly 50x cost difference. (author_report, not tested)

Worth knowing

1. The author mentions 'I made a prompt' but does not include it in the post; the reproducible content is the step-by-step guide above. 2. Exact env vars, Claude Code configuration commands, and Grok Bot routine setup UI/API are not provided; readers must consult respective documentation. 3. The criteria for 'risky' are undefined; users must set their own thresholds. 4. The post quotes another tweet (https://twitter.com/yurshevv/status/2111493684189368536), but that content is not available in the snapshot and cannot be verified. 5. This site has not performed any reproduction or validation.

Steps

  1. Get a DeepSeek API key

    Request an API key from DeepSeek. The author cites DeepSeek's own chart that Flash-1017 costs about $0.20 per DeepSWE task.

  2. Point Claude Code at DeepSeek Flash

    On the Grok Bot's computer, configure Claude Code to use DeepSeek via environment variables, with every model name set to deepseek-flash. The source does not specify exact env var names or commands; refer to Claude Code and DeepSeek documentation.

  3. Safety tip

    Safety tip: before the bot sends or submits to an external system, confirm the recipient/target and the content.

  4. Create one Grok Bot routine per job

    In Grok Bot, set up one routine per independent task. Use webhook or schedule as the wake trigger, not a 5-minute poll. The source does not provide specific UI steps or API calls.

  5. Run the build-test-fix loop

    When the routine wakes, it runs the job on DeepSeek Flash: build, test, fix. Maximum 2 attempts; if it still fails, it parks the task and notifies you.

  6. Schedule heavy runs off-peak

    Run compute-heavy tasks during nights and weekends to use DeepSeek's discounted pricing.

  7. Hold risky steps for your approval

    Any risky operation (e.g., final merge) pauses and waits for your confirmation. You may use Opus 5.5 to read the final diff (review only, not execution).

Prompt you can copy

Reference prompt · Editorial suggestion。Confirm the destination, input variables and output format before use.

This block is the prompt. Ads are never placed here.

No prompt needed — just follow the steps.

Source

Grok Bot + DeepSeek Flash 夜班：例程唤醒后在 Flash 上构建测试，最多 2 次，风险步骤等人批 · yurshevv · X · 2026-10-18T20:28:41.000Z

Comments & feedback

Tried this method? Share results, problems or improvements. Comments appear after moderation. Please avoid sensitive information.

Loading…

    Related