1
# Research Brief — Skill
2
3
Turns an open question into a sourced, skimmable brief: a search plan executed with
4
source diversity in mind, every source graded for reliability, every claim tracked
5
back to a specific source in a claim ledger, and the result written up with the
6
ledger's rigor but none of its clutter.
7
8
Inputs: `question` (required), `depth` (`quick`|`standard`|`deep`, default `standard`),
9
`audience` (optional, calibrates framing/technical depth).
10
11
## Step 1 — Decompose the question
12
13
Before searching, break the question into sub-questions that, together, answer it.
14
Example: "Should we adopt library X?" decomposes into: What does X actually do
15
differently from alternatives? What's its maintenance/community health? What are
16
known limitations/gotchas from real usage? What's the migration cost? Write these
17
down — they become the structure of the search plan and, later, the brief.
18
19
## Step 2 — Build the search plan
20
21
For each sub-question, plan 1-3 searches with different angles (don't just rerun the
22
same query worded differently):
23
- A direct query for the sub-question itself.
24
- A query aimed at finding disconfirming evidence — deliberately search for
25
  criticism/limitations/failure reports, not just supporting material. A brief that
26
  only surfaces confirming sources is not research, it's confirmation bias with
27
  citations.
28
- For time-sensitive topics, a query scoped to recent results, and note the
29
  freshness of what's found.
30
31
Target source count by depth: `quick` 3-5, `standard` 6-12, `deep` 12+ with deliberate
32
diversity across source *types* (see Step 3), not just more of the same type.
33
34
## Step 3 — Grade every source
35
36
Grade each source using `source-grading.md` before using any claim from it. Record,
37
per source: type (primary/secondary/tertiary), publication/author, date, and the grade.
38
Prefer primary sources (original docs, official announcements, the actual paper/data,
39
firsthand accounts) over secondary summaries when both are available — a summary can
40
introduce or drop nuance the brief needs.
41
42
Do not use a source's claim in the brief without recording it in the source list, even
43
if it "seems obviously true" — the ledger in Step 4 is what makes the brief checkable
44
later.
45
46
## Step 4 — Build the claim ledger
47
48
For every specific factual claim that will appear in the brief, record: the claim
49
itself (stated precisely, not vaguely), which source(s) support it, and — critically —
50
whether other sources found in Step 2 *agree*, *disagree*, or *don't address* it.
51
52
- A claim supported by only one low-graded source is weak — either find corroboration,
53
  soften the claim's confidence in the brief ("one report suggests..." vs. stated as
54
  fact), or drop it.
55
- A claim where sources genuinely disagree is not a bug to resolve by picking a side
56
  arbitrarily — the disagreement itself is often the most useful thing to report to
57
  the reader.
58
- Numbers/statistics get special scrutiny: check the original methodology when
59
  possible (sample size, date, what was actually measured) rather than repeating a
60
  number secondhand.
61
62
## Step 5 — Write the brief
63
64
Use `templates/brief.md`. Calibrate to `audience` if given — a brief for engineering
65
leadership can assume technical vocabulary; a brief for general readers should define
66
terms. Lead with the direct answer to the original `question`, then support it —
67
don't make the reader wait for a narrative buildup to find out the conclusion.
68
69
Every non-obvious claim in the prose carries an inline citation back to the source
70
list (not the raw ledger — the ledger is working material, the brief cites sources).
71
State confidence honestly: distinguish "well-established across multiple independent
72
sources" from "single source, unverified" from "sources disagree" — don't flatten
73
everything to the same confident tone.
74
75
## Step 6 — Self-check before delivering
76
77
- Does every claim in the brief trace to an entry in the claim ledger? Anything that
78
  doesn't must be either sourced or removed/flagged as the writer's own inference.
79
- Did the search plan actually look for disconfirming evidence (Step 2), and if it
80
  found any, is it represented in the brief — not just the supporting evidence?
81
- Is the direct answer to `question` stated plainly, near the top?
82
- Are dates/versions/numbers specific rather than vague ("as of [date]" not "recently")?
83
84
## Stop conditions
85
86
- Fewer than the target source count can be found after a genuine search effort
87
  (topic too obscure/new) → say so explicitly in the brief's confidence framing rather
88
  than padding with low-quality sources to hit a number.
89
- Sources are irreconcilably split on a load-bearing claim → present the disagreement
90
  and each side's basis rather than picking one arbitrarily.
91
- The question itself is unanswerable as posed (too vague, conflates two different
92
  questions) → say so and propose a decomposition rather than guessing at intent.
93