AI Usage Guides · 2026-11-02

DeepSeek Writing Templates: Weekly Reports, Retros, and Proposals

Weekly reports, project retros, and proposals are three high-frequency writing tasks that are easy to get stuck on. DeepSeek has a particular strength here: the R-series deep thinking mode is good at step-by-step reasoning under constraints, but if you pile on roleplay flourishes like "you are a senior expert" or "be creative," results are often worse than simply stating the structure, materials, and word count clearly. This post gives you three ready-to-copy templates, all following the same principle: spell out input materials, output structure, and style constraints, then let the model do the rest.

First, Know When to Turn On Deep Thinking

DeepSeek's chat modes come in two lines: the V series handles general conversation with fast responses, suitable for everyday rewriting, polishing, and formatting; the R series, with deep thinking enabled, reasons before answering, making it better for tasks that require analysis—like distilling retros from scattered notes or weighing options between proposals. You can toggle the switch near the chat box (exact location per the official site).

Practical advice: weekly reports are "organizing" tasks—regular mode is enough and saves time; retros and proposals involve trade-offs, so deep thinking is more reliable. No need to trigger deep thinking for simple tasks; the wait time grows noticeably and the payoff isn't worth it.

Also, DeepSeek handles Chinese-language context well, so just describe your needs in plain, conversational Chinese—no need to force-translate prompts into English.

Template 1: Weekly Report Generator

Dump this week's raw work log in and let it shape a structured report:

You are my writing assistant. Turn this week's raw work log into a weekly report.

[This week's log]
(Paste your unedited notes as-is)

Requirements:
1. Output three sections: Completed this week / In progress / Plans for next week;
2. Summarize each item in one sentence focused on outcomes, not process; highlight deliverables and numbers;
3. Plain professional tone, no slogans or piled-up adjectives;
4. Max 5 items per section; do not fabricate anything not in the log;
5. Keep the total under 300 words.

The key is item 4, "do not fabricate"—models tend to embellish achievements when facing gaps in the material, so forbid it explicitly.

Template 2: Project Retro

A retro requires drawing causal conclusions from process notes—a classic reasoning task. Turn on deep thinking mode:

Please help me run a project retro.

[Project background]
(Goal, timeline, and my role, in 2-3 sentences)

[Process log]
(Paste key events as a timeline: what was done, what problems came up, how they were handled)

[Final outcome]
(Actual result vs. the target, and the gap)

Output structure:
1. Outcome summary: one sentence on how it went;
2. What went right: 2-3 items, each backed by evidence from the log;
3. What went wrong: 2-3 items, distinguishing "judgment errors" from "execution issues";
4. Reusable lessons: max 3, written as concrete actions for next time—no empty talk;
5. If the log has information gaps, list "facts I should still verify."

Item 5 is an anti-hallucination design: it makes the model flag missing information instead of forcing conclusions. With deep thinking on, its cross-referencing of multiple threads tends to be more solid.

Template 3: Proposal Draft

The hard part of proposals is "having ideas but no structure." Let DeepSeek build the skeleton first, then fill it in:

Help me draft a proposal. Audience: (executive briefing / team review / client pitch—pick one).

[Problem to solve]
(Current situation and pain points, within 3 sentences)

[My initial ideas]
(Write them even if they're scattered fragments)

[Constraints]
Budget: (fill in)
Timeline: (fill in)
Fixed limitations: (e.g., existing systems, compliance requirements)

Output requirements:
1. First give the proposal skeleton: goal, approach, phased plan, resources needed, risks—item by item;
2. Then write the body, each section under 200 words;
3. Compare 2 alternative approaches with their costs, to help me decide;
4. Every risk must include a mitigation—no vague "pay closer attention."

The "compare alternatives" item works best with deep thinking—having the model explicitly weigh two routes' costs and benefits beats a single handed-down answer.

Three General Tips

1. More source material beats less. The more real content fills the placeholders, the better the output; paste raw chat logs and meeting notes and let it distill. 2. Say what NOT to do. Compared to "be professional," negative constraints like "no adjective pile-ups" and "no empty talk" get followed far better. 3. Iterate rather than restart. When the first output misses, point at the specific section and how to fix it instead of rewriting the whole prompt.

All three templates are starting points. After a few rounds, fold your team's own format requirements into the constraints section, and they'll grow into a private writing pipeline that fits you. Put DeepSeek's low API pricing to work: hand these frequent small tasks over without worry—the time saved on manual formatting more than adds up.