steer-v1: 10,758 real prompts from one engineer steering coding agents
steer-v1 is 10,758 of my own prompts to coding agents (the clean variant), recorded from Dec 2025 to Oct 2026 across Codex, Claude Code, Pi, and Codex on DeepSeek and OpenRouter. Each prompt comes with the tool calls the agent made next and how I reacted on the following turn. It contains no model-generated text, and it is licensed non-exclusively.
Superseded by steer-v2. The current release has 17,126 prompts from 589 sessions, split into 13 domain packs licensed separately: see Datasets. This page describes the earlier single release.
Most agent training data is written for the occasion. This is what steering looks like when nobody is writing for a dataset: terse prompts, typos, mid-task pivots, corrections after the agent has already acted, and sessions that run for days.
What's in it
| Measure | Value |
|---|---|
| Prompts | 10,758 in the clean variant (11,720 in the full set); median 88 characters |
| Sessions | 343, including 53 sessions of 50+ prompts; the longest has 556 |
| Agent tool calls between prompts | 62,359 across the full set (shell, patch/edit, file read/write, planning) |
| Harnesses (full set) | Codex CLI 8,314 · Claude Code 2,096 · Codex on DeepSeek 707 · Pi 443 · Codex on OpenRouter 160 |
| Model families steered | OpenAI, Anthropic, DeepSeek, Moonshot, Xiaomi, Zhipu, with exact model ids per record |
| Span | Dec 2025 → Oct 2026 |
| Domains | Agent harnesses and ports, CUDA/GPU rendering, trading systems, RL and training, LLM weight editing, phone and computer-use agents, Linux workstation setup, audio and video studios, IoT, security |
One record
Each line of steer.jsonl is one human turn:
{"session": "16-hex id", "turn": 12, "harness": "codex", "models": ["…"], "model_family": ["openai"],
"project_family": "RL & training", "month": "2026-08", "minutes_since_start": 41.5,
"prompt": "the engineer's prompt (scrubbed)", "prompt_chars": 88,
"agent_tool_calls": 6, "agent_tools": {"exec_command": 4, "apply_patch": 2},
"next_reaction": "correction | approval | neutral | session_end",
"abusive": false, "profane": false, "label_source": "keyword-heuristic"}sessions.jsonl lists each session's harness, models, project family and prompt count.
What it's good for
- Training and evaluating agents on how real users steer: short instructions, pivots, and corrections that arrive after the agent acted.
- Reaction modelling:
next_reactionis a weak signal of whether the agent's action landed. - Cross-harness comparison: one person driving five harnesses and 20+ models on the same kinds of work.
How it was cleaned
- No model text. Only my prompts, plus the names and counts of the agent's tool calls. Briefs that agents wrote and I pasted on, agent-to-agent relay messages and pasted terminal output were removed. Long pasted blocks inside prompts are replaced by a size marker.
- No client work. Sessions in client folders are excluded, and client and project code names are replaced by
[CLIENT]. - No personal or confidential material, including take-home or interview work.
- Secrets (keys, tokens, JWTs, private keys, long hex or base64 strings,
key=valuecredentials) become[SECRET]. The release was checked against every real credential value on the source machine, with zero matches. - Emails, phone-like numbers, IP addresses, home paths and my name are masked.
- Strong language is flagged per record (
abusive,profane), so you can filter it. The clean variant drops everyabusiverecord.
Limitations
next_reactioncomes from a keyword heuristic, not human labels.- Tool calls are names and counts only. The agent's arguments and outputs are not included.
- It is one person: deep and realistic, not demographically broad.
How to get it
steer-v1 is licensed non-exclusively, under the agreement attached to the listing. The source transcripts stay with me. Email me for the datasheet and a scrubbed sample.
FAQ
Does steer-v1 include model outputs?
No. It contains only the human prompts, the names and counts of the agent's tool calls, and the next human reaction. No assistant or model-generated text is included.
Is any of it synthetic?
No. Every prompt was written by me while working on real projects with coding agents between December 2025 and October 2026.
How many prompts and sessions are there?
The clean variant has 10,758 prompts across 343 sessions; the full set has 11,720 prompts. 53 sessions have 50 or more prompts, and the longest has 556.
Which agents and models does it cover?
Five harnesses (Codex CLI, Claude Code, Pi, and Codex on DeepSeek and OpenRouter) and model families from OpenAI, Anthropic, DeepSeek, Moonshot, Xiaomi and Zhipu, with exact model ids on each record.
Is the licence exclusive?
No. It is licensed non-exclusively, and the exact rights depend on the agreement attached to the listing.
License it, or have me build an environment in the same style.