Worked Example 2 — Prosumer Tool, with Loopback

Purpose: voice-fidelity reference per ADR-010, and demonstration of Step 3 sub-flows (script linter, interview logging, theme map, saturation, loopback) per SPEC-008-A and concept §4.3 + §7.

Scenario: Fictional builder Sam, who’s been building Notion-adjacent tools for years and has a thesis about a “knowledge graph for solo creators”. Sam runs through Steps 1–3 and ends up triggering a loopback when the interviews surface no costly-action signals despite saturated themes.

This example is intentionally a failed first attempt. Concept §7: “make a ‘no’ cheap.”

---

Step 1 — Builder Profile (abridged)

Background: 6 years building productivity tools. Three indie launches; one acquired ($120K), two flatlined. Active in two creator-economy Discord communities; ~600 Twitter followers in that niche.

Unfair advantages: First-name access to ~25 active solo creators in the productivity-tool niche; ran a Notion template store for 18 months and saw the buying behaviour up close.

Runway: 7 months solo, self-funded.

Rough idea: “AI-powered knowledge graph for solo creators — auto-link your notes by topic, pull related content into your weekly plan.”

Known buyers: ~10 solo creators willing to take a call. None have paid for a knowledge-graph tool specifically (they use Notion / Obsidian for free).

Step 2 — Vertical Stack v1

Layer	Content
Vertical	Creator economy
Sub-vertical	Solo creators publishing weekly long-form content
ICP	Solo creators with 1k–10k newsletter subscribers, US/EU, currently using Notion or Obsidian
Specific workflow	Weekly content planning — picking what to write next based on past notes and audience response
AI leverage point	Auto-link related notes by topic; surface “you wrote about this six months ago, audience liked it” prompts during planning

Backup Verticals:

1. Solo paid-newsletter operators (Substack/Beehiiv 1k+ paid subs) — same buyer with money in flight
2. Audio creators (podcasters with weekly schedules) — adjacent workflow

The Mini Council (not shown here for brevity) reaches 3 Pass for now, 2 In play at Step 1→2; Sam advances with the conditions noted (no costly-action signal yet; pricing undecided).

---

Step 3 — Sub-flow A: outreach + interview script generation

Sam runs /sprint-step-3 and picks “Starting fresh — generate the interview script and outreach plan.”

Gandalf produces a draft 9-question script. The Step 3 command shells out to the linter:

$ PYTHONPATH=${CLAUDE_PLUGIN_ROOT} python -m helpers.interview_validator /tmp/sprint-script-draft.txt
/tmp/sprint-script-draft.txt:1: ✓
/tmp/sprint-script-draft.txt:2: ✓
/tmp/sprint-script-draft.txt:3: hypothetical-would-you — "Would you…" speculates about future behaviour. Ask about past behaviour instead.
/tmp/sprint-script-draft.txt:4: ✓
/tmp/sprint-script-draft.txt:5: ✓
/tmp/sprint-script-draft.txt:6: opinion-would-it-be-useful — "Would it be useful…" leads to social-desirability bias.
/tmp/sprint-script-draft.txt:7: ✓
/tmp/sprint-script-draft.txt:8: ✓
/tmp/sprint-script-draft.txt:9: ✓
exit code: 1

The Step 3 command surfaces both violations and asks Sam: regenerate, override, or cancel. Sam picks regenerate. Gandalf re-asks questions 3 and 6 with the linter findings as feedback:

#	Original (flagged)	Regenerated (linter-clean)
3	“Would you use a tool that auto-links your notes?”	“Walk me through the last time you wished one of your notes connected to another. What did you actually do — searched manually, gave up, used a tag system?”
6	“Would it be useful to see related past content during your weekly plan?”	“Tell me about the last weekly plan you wrote. How did you decide what to write about that week? What did you reference?”

Re-run the linter on the rewritten questions:

$ PYTHONPATH=${CLAUDE_PLUGIN_ROOT} python -m helpers.interview_validator /tmp/sprint-script-draft-v2.txt
exit code: 0

Clean. Script locked into .sprint/sprint.md Interview Log → Outreach plan → Interview script.

Outreach plan also locked: target list of 22 named solo creators (drawn from Sam’s Discord and Twitter network), Sales Navigator filter (creator-economy occupation tags + 1k+ followers + US/EU), 5 communities (two Discord servers, the Indie Hackers forum, two paid newsletters’ chat groups), cold-outreach message ≤120 words anchored to “I noticed you wrote about [topic]; can I learn how you decide what to write next?”

---

Step 3 — Sub-flow B: interviews logged

Sam runs the cold outreach over 12 days. Holds 8 interviews. Each one logged via /sprint-step-3 → “Log a new interview”.

Aggregated .sprint/interviews/ directory (8 files). Excerpt of the log table written into .sprint/sprint.md:

#	Pseudonym	Date	Length	Costly action?	Themes
01	newsletter-writer-1	2026-05-12	38min	no	weekly-planning-friction, search-failure, tag-decay
02	newsletter-writer-2	2026-05-13	42min	no	weekly-planning-friction, audience-uncertainty, tag-decay
03	youtube-essayist-1	2026-05-14	35min	no	search-failure, audience-uncertainty, batch-burnout
04	newsletter-writer-3	2026-05-15	41min	unclear	weekly-planning-friction, search-failure, tag-decay
05	substack-paid-1	2026-05-16	50min	no	weekly-planning-friction, search-failure, audience-uncertainty
06	newsletter-writer-4	2026-05-18	33min	no	weekly-planning-friction, tag-decay
07	youtube-essayist-2	2026-05-19	40min	no	search-failure, audience-uncertainty
08	newsletter-writer-5	2026-05-20	36min	no	weekly-planning-friction, search-failure

Costly-Action Signals: 0 yes. One unclear (newsletter-writer-3 mentioned paying $19 for a Notion template that “kind of did this” but couldn’t remember the name; not a clear costly-action signal).

---

Step 3 — Sub-flow C: theme map

Sam runs /sprint-step-3 → “Update the theme map across all interviews”.

| Theme | Interviews | Buyers (pseudonyms) |
|---|---|---|
| **weekly-planning-friction** | 6 | newsletter-writer-1, -2, -3, -4, -5, substack-paid-1 |
| **search-failure** | 6 | newsletter-writer-1, -3, -4, -5, youtube-essayist-1, -2, substack-paid-1 |
| **audience-uncertainty** | 4 | newsletter-writer-2, youtube-essayist-1, -2, substack-paid-1 |
| **tag-decay** | 4 | newsletter-writer-1, -2, -3, -4 |
| batch-burnout | 1 | youtube-essayist-1 |

Three themes appear in 4+ interviews — strong clustering. But:

Costly-Action Signals: none. No buyer paid for a knowledge-graph tool. One paid $19 for a Notion template once. The pain is real; the wallet is not.

---

Step 3 — Sub-flow D: saturation check

Sam runs /sprint-step-3 → “Check saturation”. The command builds the JSON:

[
  ["weekly-planning-friction", "search-failure", "tag-decay"],
  ["weekly-planning-friction", "audience-uncertainty", "tag-decay"],
  ["search-failure", "audience-uncertainty", "batch-burnout"],
  ["weekly-planning-friction", "search-failure", "tag-decay"],
  ["weekly-planning-friction", "search-failure", "audience-uncertainty"],
  ["weekly-planning-friction", "tag-decay"],
  ["search-failure", "audience-uncertainty"],
  ["weekly-planning-friction", "search-failure"]
]

And shells out:

$ PYTHONPATH=${CLAUDE_PLUGIN_ROOT} python -m helpers.saturation --human /tmp/sprint-themes.json
Total interviews: 8
Themes per interview (new only): [3, 1, 1, 0, 0, 0, 0, 0]
Consecutive interviews with no new themes (from end): 5
Threshold: 3
Saturated: YES — last 3 interviews surfaced no new themes; consider stopping here.

Saturation reached. Themes converged at interview 4. Interviews 6–8 added literally nothing new.

But the theme map says: the pain is present, and the wallet is not.

---

Step 3 — Sub-flow E: the go/no-go decision

Sam runs /sprint-step-3 → “Conclude Step 3”.

The command surfaces:

Step 3 close-out.

8 interviews logged. 4 strong recurring themes (weekly-planning-friction, search-failure, audience-uncertainty, tag-decay). Saturation reached — last 5 interviews added no new themes.

0 costly-action signals. 1 unclear data point ($19 Notion template, recall failure). No buyer in this ICP has paid for a knowledge-graph tool.

The themes are real. The buyers are not paying for them.

Sam’s options:

1. Go — proceed to Step 4 with the validated themes
2. Loopback — try a different ICP / vertical / workflow
3. Pause

Sam picks Loopback. The reasoning, captured in the loopback log: themes are real but solo creators are not a paying buyer for this product class. The backup vertical (paid-newsletter operators with skin in the game) is the next attempt.

---

Loopback execution

Sam runs /sprint-loopback. The command asks:

• Loopback flavour: Different ICP, same vertical (paid-newsletter operators are still in the creator economy)
• Disqualified vertical: creator-economy / solo-creators-with-newsletters-1k-10k / weekly-content-planning
• Reason: “8 interviews, 0 costly-action signals despite saturated themes; ICP is not a paying buyer for this product class”
• Confirm: yes

The command applies the SPEC-005-A retain/clear partition:

Section	What happens
Builder Profile	Retained — Sam’s unfair advantages are still real
Backup Verticals	Retained — the list of alternatives remains useful
Vertical Stack	Cleared — the disqualified vertical is gone
Interview Log	Retained — but tagged as attempt 1 for traceability
Theme Map	Retained — these themes may apply to the next ICP too
Costly-Action Signals	Retained — empty in this case, but the section structure stays
Expert Framework	Retained — wasn’t populated yet, but the section structure stays
Competitor Map	Cleared
Positioning Whitespace	Cleared
Artifacts	Cleared

Frontmatter updates:

- step: 3
- last_approved_step: 3
- attempt: 1
+ step: 2
+ last_approved_step: 1
+ attempt: 2

Loopback Log entry appended:

- **attempt 1** — disqualified `creator-economy / solo-creators-with-newsletters-1k-10k / weekly-content-planning` (8 interviews, 0 costly-action signals despite saturated themes; ICP is not a paying buyer for this product class) (2026-05-21)

Progress Log entry appended:

- **Progress** — Loopback (different ICP, same vertical); disqualified `creator-economy / solo-creators-with-newsletters-1k-10k / weekly-content-planning` (2026-05-21)

---

Re-entry at Step 2 — Vertical Stack v2

Sam runs /sprint-step-2 (gate is now 1, attempt is 2). Gandalf re-engages, this time with two pieces of context Sam didn’t have on attempt 1:

1. The themes from attempt 1 (still in the Theme Map) — weekly-planning-friction and search-failure were the strongest signals
2. The lesson from attempt 1 — “the problem is real but solo creators don’t pay; pivot to ICPs with money in flight”

New Vertical Stack draft:

Layer	Content
Vertical	Creator economy
Sub-vertical	Paid-newsletter operators (changed)
ICP	Substack/Beehiiv operators with 1k+ paid subscribers, $5K+ MRR, US/EU (sharpened — has-revenue is the gate)
Specific workflow	Weekly content planning — same workflow, different buyer
AI leverage point	Same — auto-link, surface past topics with audience-response data

Sam advances. The Step 1→2 Mini Council convenes again on the new Stack, with the previous attempt’s evidence in scope. (Continued in v0.1.0 Phase 4+ when Steps 4–6 ship.)

---

What this example demonstrates

1. Linter-gated script generation works. Two violations on the first draft, regenerated to clean. The Step 3 command never surfaced violating questions to the user.
2. Theme map + saturation = stop signal. Sam ran 8 interviews, not because 8 was the magic number, but because saturation said so. Concept §6: “Saturation over headcount.”
3. Themes can be real and the buyer can still not pay. Strong theme clustering + zero costly-action signals = loopback, not “great theme map, let’s build”. Concept §4.3: “Stated problems are cheap. Revealed problems are signal.”
4. Loopback is structured, not chaotic. Builder Profile, Backup Verticals, Theme Map, and Expert Framework survived. Vertical Stack and downstream cleared. Sam re-entered Step 2 with everything she’d learned.
5. The cost of “no” was 12 days and 8 conversations. Not a year and a built product. Concept §7: “make a ‘no’ cheap.”

This example is the canonical reference for the Phase 3 tester gate.