Building Your First Agent

Step-by-step: capture expert knowledge, create instruction files, and run your first agentic workflow.

You’re building a research workflow. It ingests expert content, extracts key insights, and produces weekly briefings. About 30 minutes from start to finish.

Step 1: Capture expert knowledge

Pick 3-5 videos or podcasts from experts in your field. Run augent on each:

take_notes with url: "https://youtube.com/watch?v=example1"

Tag each transcription with a consistent topic:

tag with cache_key: "example1:tiny" and tags: ["your-topic"]

Consistent tagging is what makes the rest of the pipeline work.

Step 2: Find patterns across sources

search_memory with query: "best practices for [your topic]" and mode: "semantic"

Semantic search surfaces connections keyword search would miss. You can also ask Claude to read the notes directly and pull out recurring frameworks.

Step 3: Create instruction files

Create agents/research-agent/ in your vault with two files.

instructions.md:

# Research Agent

## Process
1. Find new transcriptions tagged [topic] from the past 7 days
2. Extract key claims, frameworks, actionable techniques, and contradictions
3. Cross-reference with previous briefings to avoid repetition
4. Write briefing: Key Insights, New Frameworks, Action Items, Open Questions
5. Save as "Weekly Briefing - [date].md"

## Style
- Direct, no filler
- Cite the source transcription for each insight
- Flag contradictions explicitly

rules.md:

# Rules
- Every insight must trace to a specific transcription. Never fabricate.
- Fewer than 2 new sources? Skip the briefing, report why.
- Under 800 words. Headers and bullets. No prose paragraphs over 3 sentences.
- Contradictions: present both positions, label the conflict.

These are plain markdown. Edit them anytime and the workflow’s behavior changes on the next run.

Step 4: Test and iterate

Tell Claude:

Read the files in agents/research-agent/ and follow the instructions.
Use my transcriptions tagged [topic] from the last 7 days.

Review the output. Fix what’s wrong:

Problem	Fix
Too verbose	Tighten word limit in rules.md
Missing citations	Add explicit rule requiring source file names
Redundant insights	Strengthen cross-reference step in instructions.md

Two or three passes gets you to a solid baseline.

Step 5: Save as a slash command

Create ~/.claude/commands/research-briefing.md:

Read the files in agents/research-agent/ and follow the instructions.
Use transcriptions tagged [topic] from the last 7 days.
Save the briefing to the vault.

Now /research-briefing runs the entire pipeline in one command.

Step 6: Keep feeding it

Each new take_notes call adds to the knowledge base. Week 4 briefings are better than week 1 because the context is deeper. Update the instruction files as your needs evolve. The knowledge compounds.

More examples

Content, copywriting, scriptwriting, and operations workflow templates.