返回 CodeWhale
SKILL.md
1 ---
2 name: v4-best-practices
3 description: Use when working with deepseek-v4-pro or deepseek-v4-flash in thinking mode on multi-step or plan-driven tasks. Provides rules to prevent stale references, unverified plan assumptions, and vague plan output.
4 ---
5
6 # V4 Best Practices
7
8 Rules for multi-step V4 thinking-mode workflows. Each rule prevents a
9 specific, observable failure class.
10
11 ## 1. Verify references before writing
12
13 Before referencing a file path, function, or type in code or plan output,
14 call `grep_files` or `read_file` to confirm it exists in the workspace.
15
16 ```
17 # Bad: edit_file path="src/config/loader.rs" (assumed from memory)
18 # Good: grep_files pattern="pub fn load_config" → confirms src/config/mod.rs:42
19 # then reference src/config/mod.rs:42
20 ```
21
22 Failure avoided: `edit_file` errors on non-existent paths; LSP diagnostics
23 on hallucinated symbols.
24
25 ## 2. Spawn a verifier sub-agent before multi-file execution
26
27 Before executing a plan that touches 3+ files, spawn a `deepseek-v4-flash`
28 sub-agent (thinking off) to read the target files and confirm path/symbol
29 assumptions still hold.
30
31 ```
32 agent type="verifier" model="deepseek-v4-flash"
33 prompt: "Read these files and confirm: [list assumptions]. Report mismatches."
34 ```
35
36 Failure avoided: multi-step edits fail partway because file structure
37 changed since the plan was drafted.
38
39 ## 3. Plan output must use confirmed path:line references
40
41 In plan-mode output, replace vague location pointers with `path:line`
42 references drawn from a prior `grep_files` result.
43
44 ```
45 # Bad: "Update the retry logic in the client module"
46 # Good: "Update retry loop at crates/tui/src/client.rs:187"
47 ```
48
49 Failure avoided: agent-mode execution cannot locate the intended edit
50 target when plan directions are imprecise.
51
51 lines MARKDOWN