use-cases/learn-a-new-concept.md +185 −0 added
1---
2name: Learn a new concept
3tagline: Turn dense source material into a clear, reviewable learning report.
4summary: Use Codex to study material such as research papers or courses, split
5 the reading across subagents, gather context, and produce a Markdown report
6 with diagrams.
7skills:
8 - token: $imagegen
9 description: Generate illustrative, non-exact visual assets when a Mermaid
10 diagram is not enough.
11bestFor:
12 - Individuals learning about an unfamiliar concept
13 - Dense source material that benefits from parallel reading, context
14 gathering, diagrams, and a written synthesis
15 - Turning a one-off reading session into a reusable Markdown report with
16 citations, glossary terms
17starterPrompt:
18 title: Analyze a Research Paper and Teach Me the Concept
19 body: >-
20 I want to learn a new concept from this research paper: [paper path or URL].
21
22
23 Please run this as a subagent workflow:
24
25 - Spawn one subagent to map the paper's problem statement, contribution,
26 method, experiments, and limitations.
27
28 - Spawn one subagent to gather prerequisite context and explain the
29 background terms I need.
30
31 - Spawn one subagent to inspect the figures, tables, notation, and any
32 claims that need careful verification.
33
34 - Wait for all subagents, reconcile disagreements, and avoid overclaiming
35 beyond the source material.
36
37
38 Final output:
39
40 - create `notes/[concept-name]-report.md`
41
42 - include an executive summary, glossary, paper walkthrough, concept map,
43 method diagram, evidence table, caveats, and open questions
44
45 - use Markdown-native Mermaid diagrams where diagrams help
46
47 - use imagegen to generate illustrative, non-exact visual assets when a
48 Markdown-native diagram is not enough
49
50 - cite paper sections, pages, figures, or tables whenever possible
51
52
53 Constraints:
54
55 - do not treat the paper as ground truth if the evidence is weak
56
57 - separate what the paper claims from your interpretation
58
59 - call out missing background, assumptions, and follow-up reading
60relatedLinks:
61 - label: Subagents
62 url: /codex/subagents
63 - label: Subagent concepts
64 url: /codex/concepts/subagents
65---
66
67## Introduction
68
69Learning a new concept from a dense paper or course requires more than just summarization. The goal is to build a working mental model: what problem it addresses, what the method actually does, which evidence supports it, what assumptions it depends on, and which parts you still need to investigate.
70
71Codex is useful here because it can automate the context gathering, and can turn complicated concepts into helpful diagrams or illustrations. This use case is also a good fit for [subagents](https://developers.openai.com/codex/concepts/subagents): one thread can read the paper for structure, another can gather prerequisite context, another can inspect figures and notation, and the main thread can reconcile the results into a report you can review later.
72
73For this use case, the final artifact should be something you can easily review: a Markdown file such as `notes/concept-report.md`, or a document of another format. It should include a summary, glossary, walkthrough, diagrams, evidence table, limitations, and open questions instead of ending with a transient chat answer.
74
75## Define the learning goal
76
77Start by naming the concept and the output you want. A narrow question makes the report more useful than a broad summary.
78
79For example:
80
81> I want to understand the main idea in this research paper, how the method works, why the experiments support or do not support the claim, and what I should read next.
82
83That scope gives Codex a concrete job. It should teach you the concept, but it should also preserve uncertainty, cite where claims came from, and separate the paper's claims from its own interpretation.
84
85## Running example: research paper analysis
86
87Suppose you want to learn about a paper about an unfamiliar model architecture. You want a report that lets you understand the concept at a glance, without having to read the whole paper.
88
89A good result might look like this:
90
91- `notes/paper-report.md` with the main explanation.
92- `notes/figures/method-flow.mmd` or an inline Mermaid diagram for the method.
93- `notes/figures/concept-map.mmd` or a small SVG that shows how the prerequisite ideas relate.
94- An evidence table that maps claims to paper sections, pages, figures, or tables.
95- A list of follow-up readings and unresolved questions.
96
97The point is to make the learning process more systematic and to leave behind a durable artifact.
98
99## Split the work across subagents
100
101Subagents work best when each one has a bounded job and a clear return format. Ask Codex to spawn them explicitly; Codex does not need to use subagents for every reading task, but parallel exploration helps when the paper is long or conceptually dense.
102
103For a research paper, a practical split is:
104
105- **Paper map:** Extract the problem statement, contribution, method, experiments, limitations, and claimed results.
106- **Prerequisite context:** Explain background terms, related concepts, and any prior work the paper assumes.
107- **Notation and figures:** Walk through equations, algorithms, diagrams, figures, and tables.
108- **Skeptical reviewer:** Check whether the evidence supports the claims, list caveats, and identify missing baselines or unclear assumptions.
109
110The main agent should wait for those subagents, compare their answers, and resolve contradictions. Codex will then synthesize the results into a coherent report.
111
112## Gather additional context deliberately
113
114When the paper assumes background you do not have, ask Codex to gather context from approved sources. That might mean local notes, a bibliography folder, linked papers, web search if enabled, or a connected knowledge base.
115
116If you're learning about an internal concept, you can connect multiple sources with [plugins](https://developers.openai.com/codex/plugins) to create a knowledge base.
117
118Keep this step bounded. Tell Codex what counts as a reliable source and what the final report should do with external context:
119
120- Define prerequisite terms in a glossary.
121- Add a short "background you need first" section.
122- Link follow-up readings separately from the paper's own claims.
123- Mark claims that come from outside the paper.
124
125## Generate diagrams for the report
126
127Diagrams are often the fastest way to check whether you really understand a concept. For a Markdown report, ask Codex for diagrams that stay close to the source material and are easy to revise.
128
129Good defaults include:
130
131- A concept map that shows prerequisite ideas and how they connect.
132- A method flow diagram that traces inputs, transformations, model components, and outputs.
133- An experiment map that connects datasets, metrics, baselines, and reported claims.
134- A limitations diagram that separates assumptions, failure modes, and open questions.
135
136For Markdown-first reports, ask for Mermaid when the destination supports it, or a small checked-in SVG/PNG asset when it does not. Ask Codex to use the imagegen system skill, which comes with Codex by default, only when you need an illustrative, non-exact visual or something that doesn't fit in a Markdown-native diagram.
137
138## Write the Markdown report
139
140Ask Codex to make the report self-contained enough that you can return to it later. A useful structure is:
141
1421. Executive summary.
1432. What to know before reading.
1443. Key terms and notation.
1454. Paper walkthrough.
1465. Method diagram.
1476. Evidence table.
1487. What the paper does not prove.
1498. Open questions and follow-up reading.
150
151The report should include source references wherever possible. For a PDF, ask for page, section, figure, or table references. If Codex cannot extract exact page references, it should say that and use section or heading references instead.
152
153## Use the report as a study loop
154
155The first report is a starting point. After reading it, ask follow-up questions and have Codex revise the artifact.
156
157Useful follow-ups include:
158
159- Which part of this method should I understand first?
160- What is the simplest toy example that demonstrates the core idea?
161- Which figure is doing the most work in the paper's argument?
162- Which claim is weakest or least supported?
163- What should I read next if I want to implement this?
164
165When the concept requires experimentation, ask Codex to add a small notebook or script that recreates a toy version of the idea. Keep that scratch work linked from the Markdown report so the explanation and the experiment stay together.
166
167Example prompt:
168
169## Skills to consider
170
171Use skills only when they match the artifact you want:
172
173- `$jupyter-notebook` for toy examples, charts, or lightweight reproductions that should be runnable.
174- `$imagegen` for illustrative visual assets that do not need to be exact technical diagrams.
175- `$slides` when you want to turn the report into a presentation after the learning pass is done.
176
177For most paper-analysis reports, Markdown-native diagrams or simple SVG files are better defaults than a generated bitmap. They are easier to diff, review, and update when your understanding changes.
178
179## Suggested prompts
180
181**Create the Report Outline First**
182
183**Build Diagrams for the Concept**
184
185**Turn the Report Into a Study Plan**