DeepSeek Harness First Task: Rewrite the Quickstart Prompt
The DeepSeek Harness quickstart gives exactly one first task: “Summarize this repository and identify its main packages.” It is short, but it is not random: it fixes the object, action, scope, and output, so the agent can tell when it is done (source). A good first task also separates configuration problems from model-quality problems.
DeepSeek Harness: What the official example does
The sentence does three things: it points at the current repository, asks for a summary, and asks for the main packages. It does not edit files and does not run tests. That makes it a read-only task, which is the right shape for a first run: a failure points at configuration or reading, not at edit quality. (source)
The same docs say the agent can read and edit workspace files, run commands, delegate work, and maintain a plan; operations that need approval ask you first (source). The first task could be much more complex, but it should not be, because complexity mixes variables.
Treat the example as a template, not a magic sentence. Its value is the structure: where to look, what to do, what to focus on, what to return. You can translate it or replace the repository without losing that structure.
If the English feels awkward, paraphrase rather than memorize. Replace “this repository” with the project name, “summarize” with your verb, and “main packages” with your scope. Do not drop the qualifiers.
DeepSeek Harness: Split one sentence into four parts
Object, action, scope, and output are a practical description check, not an official taxonomy. The mapping looks like this (source):
| Part | In the official example | What it does | Missing means |
|---|---|---|---|
| Object | this repository | Tells the agent what to touch | It may answer generically or read the wrong place |
| Action | Summarize | Tells it what to do | It may list files without a conclusion |
| Scope | main packages | Limits the focus | It may explain every file |
| Output | identify ... packages | Defines completion | It may stop halfway or keep asking |
You do not need four sentences. The official example fits all four into one, using qualifiers. A Chinese task can do the same; just keep all four parts.
A quick check: after writing one sentence, ask “what, do what, focus on what, return what.” Four answers mean the task is clear enough; a missing answer is where the agent will stall.
DeepSeek Harness: What missing one part looks like
Different missing parts fail differently, so the symptom can tell you what to fix. Four common patterns (source):
- Missing object — the reply cites no concrete file and reads like generic documentation, or the agent asks which directory to inspect.
- Missing action — the reply lists directories and files without a summary, comparison, or conclusion.
- Missing scope — the reply is long and covers dependencies, scripts, and docs; the point is buried.
- Missing output — the reply stops midway or keeps asking “is this okay?” because it has no completion condition.
Use the symptom backward: do not switch models first; check which part is missing. Many “the model is bad” cases are “the task did not say.”
Two missing parts at once: restore object and output first. The object decides where it looks; the output decides when it stops. Action and scope can be narrowed after the first reply.
Four parts are a floor, not a ceiling. You can add “do not modify files” or “only inspect the top two directories.” For a first task, keep constraints few; every extra constraint is another possible cause of failure.
DeepSeek Harness: Three reusable rewrites
The same four-part frame works for different first tasks. All three below stay read-only (source):
- Repository summary — “Summarize what this repository is for, list the main packages and what each owns, and do not modify files.”
- Object: current repo; action: summarize; scope: main packages; output: purpose plus package roles.
- Single-file explanation — “Explain the entry logic in
src/index.ts, list the three key modules it calls in execution order, and do not change code.”- Object: one file; action: explain; scope: entry logic and three modules; output: ordered explanation.
- Failing-test location — “Find the failing tests, list each failing case, the error summary, and the most likely source file; do not fix anything yet.”
- Object: tests; action: locate; scope: failures and source files; output: list plus judgment.
The shared move is “read first.” That is not timidity; it is a path check. If the agent can read files, produce a conclusion, and follow your output shape, the model and workspace are working. Then ask for edits.
Do not put multiple goals in the first task. “Summarize the repo and fix all tests” mixes reading and writing; when it fails, you cannot tell which half broke.
DeepSeek Harness: Templates and read-before-edit
English and Chinese templates share the same structure; only the wording changes. Two templates (source):
English:
Read {object} in the current workspace, {action}, focusing on {scope}, and return {output}. Do not modify any files yet.
Chinese:
请阅读当前工作区里的 {对象},{动作},重点看 {范围},最后用 {产出形式} 给出结果;先不要修改任何文件。
Both keep “do not modify yet.” The line is not a limit on the agent’s ability; it narrows the first task to reading, so a failure tells you whether reading or editing broke.
Why read before edit? The failure surface is smaller: a missing workspace, a bad model key, or an unreadable file shows up in the summary. An edit adds “was the change correct?” on top. Remove the first layer before adding the second.
The summary must also be checkable. If the reply never mentions the file, directory, or package you named, do not move to edits. Ask for citations or narrow the scope to one file. A checkable summary is more useful than a polished overview.
DeepSeek Harness: How to tell the first task was written well
The criterion is coverage of the four parts, not whether the reply looks smart. Four checks (source):
- Object — the reply cites the repository, file, or directory you named.
- Action — it summarizes, explains, or locates, rather than only listing names.
- Scope — it stays on the focus you named instead of covering everything.
- Output — the shape matches what you asked for: list, table, or ordered explanation.
All four passing means the task is well written and the model, workspace, and execution path work. If one fails, change the task description before changing the model; it is the cheapest variable.
Once the read-only task passes, add edits, commands, or plugins. The four-gate view is in DeepSeek Harness quickstart: four gates; the path from startup to the first message is in The first five minutes with dsh web.
FAQ
DeepSeek Harness supports that. The DeepSeek Harness quickstart example, Summarize this repository and identify its main packages, already contains object, action, scope, and output; keep those four when translating.
In DeepSeek Harness, a read-only task separates whether the model, workspace, and file access work from whether edits are correct. DeepSeek Harness can edit files, but a first task should prove the path before changing anything.
In DeepSeek Harness, the reply tends to be generic, cite no concrete files, or ask what you want. Without an object and output constraint, the DeepSeek Harness agent cannot tell when it is done.
DeepSeek Harness does not work that way. One sentence is enough when it names the object, action, scope, and output. A longer prompt that mixes several goals makes it harder to tell which part failed.
Related Terms
- first task
- The first task is the first request in a new DeepSeek Harness session; it checks the model, workspace, and execution path together.— DeepSeek Harness docs - Using the Web UI
- task four
- Task four is a description check made of object, action, scope, and output, used to decide whether a DeepSeek Harness task is specific enough.— DeepSeek Harness docs - Using the Web UI
- read-only task
- A read-only task asks a DeepSeek Harness agent to read, summarize, or explain without modifying files; it is the safest first task.— DeepSeek Harness docs - Using the Web UI
- output constraint
- An output constraint defines the result shape, such as a package list or file paths, so a DeepSeek Harness agent knows when to stop.— DeepSeek Harness docs - Using the Web UI
Sources
- DeepSeek Harness docs - Using the Web UI· deepseek-harness
- DeepSeek Harness docs - Configuring models· deepseek-harness