# Discovery Agent Stack — Starter Pack

Companion to [Build a Discovery Agent Stack](https://falkster.com/blog/build-discovery-agent-stack). Install commands, three ready-to-paste prompts for the three core recipes, and the daily rhythm.

Part 2 of 5 in the [PM Agent Stack series](https://falkster.com/blog/pm-agent-stack-overview).

---

## Install (60 to 90 minutes on a clean machine)

```bash
# 1. Base — install if not present
npm install -g @anthropic-ai/claude-code

# 2. Official skills + customer-call-notes skill
#    Add anthropics/skills via Claude Code's plugin marketplace
#    Add the customer-call-notes skill (lives in falkster.ai cowork mode setup
#    and on GitHub)

# 3. GitHub MCP
#    Configure github/github-mcp-server with a PAT (repo:read scope)

# 4. Superpowers
/plugin marketplace add obra/superpowers-marketplace

# 5. claude-mem (memory)
#    Install via the marketplace or follow the README at
#    github.com/thedotmack/claude-mem

# 6. Playwright MCP (browser automation)
#    Install microsoft/playwright-mcp following its README

# 7. Subagents — pick THREE: research synthesizer, devil's advocate,
#    customer-empathy critic. Install via wshobson/agents marketplace.
```

Do not install all seven at once. Install the first three (Claude Code, official skills + customer-call-notes, github-mcp). Use them for a week. Then add the next four if the friction is still there.

---

## Recipe 1 — Real-time interview synthesis

**During the call** (after the customer has been talking for 10 to 15 minutes), paste the running transcript into Claude Code:

```
Summarize the patterns you see so far in this customer interview transcript.
Identify problem statements (vs. solution statements), pull quotable evidence,
and propose three follow-up questions to ask in the second half of the call.

Transcript:
[paste running transcript]
```

**After the call**, paste the full transcript:

```
Run the customer-call-notes skill on this transcript.
Then compare to the last five customer-call-notes outputs in my notes folder
(research/calls/). Report: what is new, what is repeating, what disconfirms
prior themes.

Transcript:
[paste full transcript]
```

Save the output to `research/calls/{date}-{customer-initials}.md`.

---

## Recipe 2 — Weekly theme tracking on autopilot

**Setup folder structure once:**

```
research/
  calls/        ← {date}-{customer}.md files
  tickets/      ← {date}-tickets-batch.md
  reviews/      ← {date}-reviews-batch.md
  themes.md     ← living document of current themes
  MEMORY.md     ← index pointing claude-mem at the folders
```

**Friday morning prompt** (run weekly, manual or via scheduled hook):

```
Read all files in research/ added since {LAST_RUN_DATE}.
Compare to the themes in research/themes.md.

Report:
1. Themes that strengthened (with evidence)
2. Themes that weakened (with evidence)
3. New themes not yet in themes.md
4. Potential weak signals worth probing in next week's interviews

Then update research/themes.md with the deltas.
Use a Sources section pointing to the specific files that produced each finding.
```

The "potential weak signals" section is consistently the most valuable. It surfaces things that wouldn't have crossed the threshold of any single interview.

---

## Recipe 3 — Signal scanning across sources

**Setup once.** Pick 3 to 5 public sources that talk about your product/category. For each, write a one-paragraph "what to look for" prompt naming the products, categories, and signal types.

**Daily Playwright job** (scheduled at 7 a.m. or run on demand):

```
For each of the following sources, fetch new content posted since {YESTERDAY}:
- {source 1 URL} — {what to look for}
- {source 2 URL} — {what to look for}
- {source 3 URL} — {what to look for}

For each source, report new themes, surprising signals, and direct quotes
worth flagging. Skip sources where nothing notable appeared.

Output as a one-page Slack message I can read in under 90 seconds.
```

**Tuning expectations:** week one will produce noise. Two of five sources will produce real signal by week three. Tune the prompts. Retire sources that consistently produce noise.

---

## Daily/weekly rhythm

| When | What |
|---|---|
| During each customer call | Run recipe 1's "during the call" prompt at the 10-minute mark |
| Right after each call | Run recipe 1's "after the call" prompt; save to research/calls/ |
| Daily (scheduled, 7 a.m.) | Run recipe 3's signal scan; check the Slack digest |
| Friday morning | Run recipe 2's weekly theme tracker; update research/themes.md |
| Monthly | Skim awesome-claude-code for new tools; tune the prompts |

---

## Critical limitations

- The agent only sees the transcripts and sources YOU give it access to. Anything outside that scope is invisible.
- Treat customer transcripts the way your company's data handling policy says to treat them.
- Public scraping respects site terms of service: one fetch per source per day, never republish raw content.
- The agent is not a replacement for direct customer contact. Continuous discovery still requires you in the room.

**Sources:** [Discovery agent stack post](https://falkster.com/blog/build-discovery-agent-stack) · [Continuous discovery on autopilot](https://falkster.com/os/continuous-discovery-autopilot) · [Interview guide](https://falkster.com/os/interview-guide)