How it works

Find the stage. Make one move. Measure until you know.

Gamify It turns three books and the research on experiments into one loop your assistant can run with you, week after week.

The loop

Seven phases, one constraint at a time.

Framing and locating run once, and again when the bottleneck moves. Diagnose to learn repeats for each constraint.

00Product Passport
Frame
What the product is, the action that creates value, how often a devoted user should return, and what you can change. Under 10 minutes.
01Stage Card
Locate
Your maturity and the one binding cell, with a confidence level, a runner-up and the three facts that would change the answer.
02Up to three root causes
Diagnose
Only the stage that binds, with only the lenses it needs: ability first, then motivation, then the cues at each decision point.
03Opportunity Board
Decide
One Next Move and two alternates, ranked by Size × Odds × Fit ÷ Cost, plus the no-regret fixes and the Not-now list.
04Experiment Card and a PR
Build
In Claude Code and Codex, your assistant builds it behind a flag and opens a pull request. Elsewhere, setup steps for your own tools.
05Health checks
Measure
Exposure and sample-ratio checks while it runs. No peeking at the effect before the pre-registered horizon.
06Readout and Learning Log
Learn and re-stage
Ship, iterate, kill or Inconclusive under the rule you set up front. Durability is checked weeks later.
Then, back in
- To Diagnose. Same bottleneck, next move.
- To Locate. A re-stage trigger fired: the bottleneck moved.
- To Frame. The business model or product type changed.

A full turn takes as long as the readout needs, typically 2 to 10 weeks. Every waiting week gets one concrete, 15-minute task in the Monday brief. We never promise an experiment every week.

The Stage Card

Readable in sixty seconds. Expert detail is one question away.

Every host shows the same card, laid out by the server, with or without our skills installed.

Stage Card · Crewbook

Fictional, illustrative

Shift-scheduling SaaS for independent restaurants, US · 2026-10-23

Where you are

Early fit (M2), confidence medium.

Restaurants with 10+ staff keep using it (31% still active at week 8); under 10 staff fall to about 4%. Unknown: the PMF survey hasn’t run yet.

Readiness

Your key action is tracked; what brings each session in is not. About 90 sign-ups a week.

Your bottleneck

Start. Only 32% of new restaurants publish a schedule that 3+ staff confirm within 7 days. Those that do are about 6× likelier to be active at week 8 (54% vs 9%; a correlation, not yet causation).

Non-obvious

Your SMS reminders go only to the manager, so staff never see the value that makes a manager come back. The trigger score is capped at 2: no trigger reaches the people whose confirmations are the reward.

Not now

Paid ads (fix the setup leak first) · annual-plan discounts (there’s no habit to protect yet) · points or badges (first steer new restaurants onto the path your retained restaurants took).

Next move

“First schedule in 10 minutes”: import staff from a CSV, generate a draft week, send confirm-by-text to staff, and ask for availability rules after.

How we’ll know

Primary: published + 3 confirmations in 7 days. Readout in about 6 weeks (a bundled change). Guardrails: 7-day trial cancels; week-2 retention.

If we’re wrong

The real bottleneck may be Value for restaurants under 10 staff.

To be more sure

(1) Run the one-question PMF survey with active restaurants. (2) Export PostHog sign-ups by week. (3) Record what brings each session in.

Sources

HK-INV19, HK-INV01, HK-ACT32, HK-TRG08 (Hooked ch. 2, 3, 5) · GT#43 and the First Major Win-State (Actionable Gamification)

A stage, with its confidence
Maturity comes from product-market-fit signals only, never traffic. The card says how sure it is and what is still unknown.
Claims capped by your tracking
What your analytics can support limits what the card may claim, and sets whether a cheap measurement lane runs alongside.
One cell, not forty ideas
Ten Constraint Rules pick the single binding cell on the Stage Map. Everything else waits.
Specific to your product
Server-side contradiction checks surface one insight about your product, not a generic tip from a book.
What to skip, and why
Good advice at the wrong stage is bad advice. The card names the tempting moves that are out of season for you.
One move, every factor shown
Moves are ranked by Size × Odds × Fit ÷ Cost, and you can argue with each factor.
Sized to your traffic
The test is sized before you build. If your traffic can’t detect the change, you get a bigger swing or the low-traffic loop instead of a micro-tweak.
A named runner-up
If the stage call is wrong, the card says where the work would be, so one wrong call can’t block the right work.
The three facts that matter
The questions worth the most, in order. Answering them raises confidence; the card never pretends to know.
Cited, not invented
Book and technique IDs sit on one line at the end, paraphrased with page references. Full citations are one question away.

Evidence levels

Every input says how we know it.

Confidence is high only when the key inputs are measured and agree. Scores are capped by their evidence, measured numbers older than 90 days count one level lower, and conflicting signals become questions instead of averages.

What we don’t claim

The Hook Model is a practical tool, not a validated or complete theory. Octalysis is a practitioner framework, so its predictions are tested in your product. Left brain and right brain is a metaphor.

Evidence levels E0 to E4
Level	What it means	Example
Evidence level E0, Unknown: Not investigated yet; an assumption or archetype default.	Not investigated yet; an assumption or archetype default.	A default for your product type, not yet checked.
Evidence level E1, Stated: Stated by the team or found in desk research.	Stated by the team or found in desk research.	“Most users come back daily,” says the founder.
Evidence level E2, Observed: Seen in the code, the screens or the live product.	Seen in the code, the screens or the live product.	Onboarding has 7 required steps, seen in the code.
Evidence level E3, Measured: Measured in aggregate data or exports.	Measured in aggregate data or exports.	Week-8 retention is 31% in segment A, with n reported.
Evidence level E4, Tested: Tested with an experiment in this product.	Tested with an experiment in this product.	An experiment in your own product, read out with a confidence interval.

Grades for every source

Research is graded A to D, with X for contested findings. Case numbers carry two grades: one for the number’s source, and one for how well we know the book mechanism explains it.

Evidence grade A. Robust evidence: replicated, peer-reviewed research.
Evidence grade B. Moderate evidence: solid studies with some limits.
Evidence grade C. Limited evidence: few or small studies.
Evidence grade D. Practitioner framework or anecdote: plausible, untested.
Evidence grade X. Contested: the evidence conflicts or failed to replicate.

The honesty contract

We never promise a lift.
Every number carries its source, its date and its grade.
Readouts come with confidence intervals, and “Inconclusive” is a legitimate outcome.
Thresholds like a 40% PMF score are labeled heuristics, not laws.
Frameworks are labeled for what they are: the Hook Model is a practical tool, and Octalysis is a practitioner framework.

A session, start to readout

Watch one turn of the loop.

Crewbook is a fictional scheduling app for restaurants with about 90 sign-ups a week. The session runs in Claude Code: three questions, a read-only repo scan, a Stage Card, one move, honest sizing, a pull request and, six weeks later, the readout.

The honest case: Fieldnote

Fictional, illustrative

Most early products are smaller. Fieldnote, a photo-proofing tool with about 25 sign-ups a week, would need about 22 weeks for an A/B test. So it runs the low-traffic loop: an 8-week pre-registered window with a 20% holdback and a 5-user test. The readout: +8.8 points, with an interval from about −7 to +25. Decision: Inconclusive, logged as a directional learning.

“Inconclusive” is a legitimate answer. We would rather say so.

Claude Code · crewbook

Fictional, illustrative

Founder: I want more restaurants paying. What should I fix first?
Tool calls: get_statusnext_questions
No saved products yet
Assistant: I’ll read your repo (read-only) in a minute. Three quick questions first: how do restaurants pay, how many sign up each week, and what does a restaurant do in Crewbook on a good week?
Founder: $39 per location a month after a 14-day trial, card upfront. About 90 sign-ups a week. On a good week the manager publishes next week’s schedule and the staff confirm their shifts.
Tool calls: get_inspection_plan
Repo mode · read-only scan
Assistant: From the code:
- Onboarding has 9 screens before the schedule grid appears.
- SMS reminders go only to the manager. Staff get an email asking them to download an app.
- PostHog tracks sign-ups and published schedules, but not onboarding steps, staff confirmations or what brings each session in.
- Availability rules are stored, but nothing uses them to draft next week.
Tool calls: locate_stage
Stage Card v0 · low confidence
Assistant: A first, provisional read: your bottleneck looks like Start, and the non-obvious part is that your reminders never reach staff. Two read-only queries through your Supabase connector would make this much sharper. OK?
Founder: Go ahead.
Tool calls: your Supabase connector
12,400 sign-ups since launch: production, not seed data · only aggregates go to Gamify It
Assistant: From the cohort query (12 weeks):
- 32% of new restaurants publish a schedule that 3+ staff confirm within 7 days.
- Those that do are 54% active at week 8, vs 9% for the rest. That’s a correlation, for now.
- Restaurants with 10+ staff: 31% still active at week 8. Under 10 staff: 4%.
- Trial to paid is 41% past that threshold, and 7% for the rest.
Tool calls: locate_stagescore_hookscore_drivesmap_decision_points
Binding stage only
Stage Card · Crewbook
Where you are
Early fit (M2), medium confidence
Bottleneck
Start: 68% never get a schedule published and confirmed in week 1
Non-obvious
Reminders never reach staff, so the manager never sees the confirmations
Not now
Paid ads · annual discounts · points or badges
If we’re wrong
Value, for restaurants under 10 staff
Tool calls: rank_opportunities
90 eligible a week · 3-week effort budget

Opportunity Board
Option	Priority	Detectability
Next Move: “First schedule in 10 minutes” (bundle)	0.7	Ready, ~6 weeks
Easy: confirm-by-text to staff only	1.3	Underpowered alone (~20 weeks)
Bigger swing: concierge setup for 10+ staff	0.2	Directional only

Assistant: The easy option scores higher on its own, but it can’t be read at your volume, so it rides inside the bundle: import staff from a CSV, draft a week from a template, text staff a confirm link, show live confirmations, and only then ask for availability rules.
Tool calls: design_experiment
Baseline 32% · 90 a week · 7-day metric
Assistant: Honest sizing:
- A 20% lift on 32% needs about 850 restaurants per arm: about 20 weeks. Not worth it.
- The bundle is designed to detect a 40% lift: about 213 per arm, about 6 weeks including the 7-day metric.
- Prior range 0 to +40%; P(win) about 25%, uncalibrated. Not a forecast.
- Guardrails: 7-day trial cancels and week-2 retention.
Founder: Do it.
Tool calls: gamifyit-build (skill)save_session
Behind the flag onboarding_first_schedule · PR #58
Assistant: PR #58 is ready for review, and EXP-007 is saved with the readout due Monday, December 14. This week’s task: send a one-question PMF survey to your active restaurants.
Two weeks after launch
The split is healthy (88 vs 91 restaurants), but only about 41% of the sample is in, so the effect isn’t read yet. The survey came back at 47% “very disappointed” among 10+ staff restaurants: stage confidence is now high.
Tool calls: compute_metricssave_experiment
Readout · December 14
Assistant: Readout under your pre-registered rule:
- Control 68 of 212 (32.1%); variant 99 of 214 (46.3%). Effect +14.2 points, 95% CI +5.0 to +23.4.
- That’s above the prior range, so I checked the logs first: exposure and events reconcile in both arms.
- Shrunk estimate: about +9 points. A 10% holdback re-read decides the durable number.
- Trial cancels 4.7% vs 4.2%: no significant harm, but that guardrail is not powered.
- Decision: Ship, with a 60-day holdback. Next binding cell: Return.

Readout · step 24 of 24

Who does what

Nothing leaves your side but a summary.

Gamify It never receives your source code, raw event rows or personal data about your users. It rejects payloads that look like them, and diagnostic tools store nothing until you save.

Your assistant: the eyes and hands

Reads your code, pages or screenshots, on your side
Turns what it finds into a short, typed evidence summary
Runs read-only queries you approve, and sends aggregates only
Builds the change behind a flag when you ask

Gamify It: the brain and the memory

A deterministic engine, with no AI model on our side: same input, same answer
The knowledge base, rubrics, evidence caps and Constraint Rules
Sample sizes, readouts and confidence intervals
Your history: every Stage Card, experiment and learning, kept only when you save

Your product has its own version of this.

Find the stage you're stuck in, then make one move. Connecting Gamify It is free; Pro starts it.

Add Gamify It to your assistant Take the 2-minute check-up

Find the stage. Make one move. Measure until you know.

Frame

Locate

Diagnose

Decide

Build

Measure

Learn and re-stage

Grades for every source

The honesty contract

Your assistant: the eyes and hands

Gamify It: the brain and the memory

Your product has its own version of this.