The DeepSeek Harness guide, from zero to advanced

1. Create a dedicated folder for the job

   For example D:\work\quarterly-report. Put in it only files the agent may change.

2. Turn on version control

   git init plus one initial commit is enough. Every change then becomes comparable or revertible.

3. Point the agent at that folder

   With the CLI, run the command from inside that folder. With the web interface or desktop app, pick it via Choose workspace.

Output quality is driven by how you ask. These four patterns almost always improve results: state the goal, the constraints, the output shape, and the verification.

• Name the folders or files that may be touched — and those that may not.
• Specify the output shape: a table, a Word file, a change list or a short summary.
• Ask the agent to verify itself, for example by running tests or reopening the result.
• For risky work, ask for a plan first without changing any files.
• Split large jobs into rounds you can review.

Tujuan: rapikan struktur folder dokumen/ agar konsisten.
Batasan: jangan ubah isi berkas, jangan hapus apa pun, hanya pindahkan dan ganti nama.
Keluaran: tabel berisi 'lama -> baru' plus skrip yang bisa saya jalankan sendiri.
Verifikasi: setelah selesai, daftar berkas yang tersisa di dokumen/ dan pastikan jumlahnya sama.

This four-line pattern works for almost any task.

A skill is a markdown file holding your procedure. Written once, the agent follows it every time that task appears — no re-explaining. Skills can live inside a project (project-scoped) or in your user folder (all projects).

---
name: format-laporan
description: Standar penulisan laporan internal perusahaan.
---

# Aturan
- Judul maksimal 60 karakter, tanpa singkatan.
- Wajib ada ringkasan eksekutif maksimal 150 kata.
- Semua angka menyertakan satuan dan periode.
- Simpan hasil ke laporan/ dengan nama YYYY-MM-judul.docx.

Skill names use lowercase and hyphens (kebab-case).

When work grows large — dozens of files, many sources, several viewpoints — do not pile it all into one conversation. Split it across agents. The simplest route is asking for a subagent; the most powerful is writing a workflow script.

// workflow: audit semua berkas panduan secara paralel
const files = args.files;              // daftar berkas dari argumen
phase("Audit berkas");
const results = await pipeline(
  files,
  (file) => agent(`Periksa ${file}: cari tautan mati dan klaim usang.`, { label: file }),
  (finding, file) => agent(`Verifikasi temuan pada ${file}: ${finding}`, { label: "verifikasi " + file })
);
return results.filter(Boolean);

A workflow script takes a list of items, runs each through stages, and returns structured results.