lean-experiment-kit
CLAUDE.md0 starsChanged 19 days agoMITView on GitHub
This repo runs business-idea experiments with the Testing Business Ideas method (David J. Bland and Alexander Osterwalder). One founder, small steps, fast loops. The founder makes the decisions; Claude runs the loop and keeps the record. Settings live in kit.config.json: the business name, time zone (timezone, an IANA name such as Europe/London, and utc_offset, such as +01:00), the weekend days, the day the week starts, the GitHub link and the dashboard link. Use the configured offset in every date the…
In this repository6 other files it gives its agents.
- get-startedSkill
- log-experimentSkill
- new-experimentSkill
- pull-evidenceSkill
- sync-dashboardSkill
- tourSkill
The file
We split this CLAUDE.md into the rules it gives an agent
7 rules. About 0 are common, 3 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 12
If kit.config.json's business ends with (example), the kit has not been set up yet and everything in it is the made-up Bean Box example.
- Distinctive: line 22
Learn: read the evidence honestly
- Distinctive: line 24
Manage: keep the board and the dashboard current at all times
- Very specific: line 50
Every test-card.md starts with a YAML block: id, title, segment, theme, method, boxes, status, decision, bar, result, effort, evidence, …
- Very specific: line 77
Personal data stays in CSV files only: Never an email address, phone number or name list in a .md file: every .md syncs to the dashboard.
- Distinctive: line 85
One hypothesis at a time: After a Pivot or Kill, propose exactly one next hypothesis.
- Very specific: line 95
Write the method as #NN Name exactly as in METHODS.md.
≈ “Common” is an estimate: at least 1 in 5 of the 27,641 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
CLAUDE.md — the experiment loop
This repo runs business-idea experiments with the Testing Business Ideas method (David J. Bland and Alexander Osterwalder). One founder, small steps, fast loops. The founder makes the decisions; Claude runs the loop and keeps the record.
Settings live in kit.config.json: the business name, time zone (timezone, an IANA name such as
Europe/London, and utc_offset, such as +01:00), the weekend days, the day the week starts, the
GitHub link and the dashboard link. Use the configured offset in every date the cards hold.
First run
If kit.config.json's business ends with (example), the kit has not been set up yet and
everything in it is the made-up Bean Box example. On the founder's first message, whatever it says,
tell them so in one line and offer two starts: /tour (see how the loop works on the example, about
five minutes) or /get-started (answer a few questions; Claude sets up their own business, about fifteen
minutes). Do not run /new-experiment, /pull-evidence or /log-experiment on the example
business unless they ask to try one on it.
The loop (repeats every time)
- Hypothesize: state a belief as "We believe that…"
- Experiment: design and run the cheapest test that gives real evidence
- Learn: read the evidence honestly
- Decide: persevere ✅ / pivot 🔄 / kill ❌
- Manage: keep the board and the dashboard current at all times
Folder layout
kit.config.json ← business name, time zone, weekend, links
customer-segments.md ← who the customers are (Seg 1, Seg 2 …)
experiments/
PROGRESS.md ← the board, GENERATED from the card headers by dashboard/build_data.py.
Never hand-edit it; edit the card.
BOARD-HISTORY.md ← corrections to the board, word for word
METHODS.md ← the 44 method names and numbers, for naming a test
_template/ ← test-card.md (with the header) and learning-card.md to copy
canvas/VP-A.md ← the Value Proposition Canvas: a header and one table, one row per box,
with Matters (1–3) and Evidence (0–3) scores; versioned in place
canvas/LC-A.md ← the Lean Canvas: the whole business model in nine blocks, the same table
and scores, two-letter IDs (CH2, RS1). VP-A details its Problem and Solution
NNNN-<slug>/ ← one folder per experiment: test-card.md (its YAML header is the record),
learning-card.md, run-log.md, daily-evidence-log.md, contact lists (CSV only)
dashboard/ ← build_data.py (cards → board and bundle), build_dash_files.py (bundle → the
dashboard's data files), dash-index.html (the dashboard page), check_sync.py
(the Stop hook) and mark_synced.py
.claude/skills/ ← the procedures below, one slash command each
docs/WALKTHROUGH.md ← one experiment from idea to decision, for a founder reading on GitHubThe card header is the record
Every test-card.md starts with a YAML block: id, title, segment, theme, method, boxes, status,
decision, bar, result, effort, evidence, note, plus window and date fields (see
_template/test-card.md). result is the whole read in one sentence. Beside it, six fields say which
of its numbers the dashboard draws as the scoreboard: read_at (required while a card is running),
score_now, score_target, score_label, score_parts and score_caveat (/pull-evidence step 4
defines them). build_data.py validates the header (status and decision values, method names against
METHODS.md, box IDs against the two canvases: one letter for VP-A, two for LC-A) and refuses to
build on a PROBLEM: line. One fact lives in one place: status and result in the header, the reading
in the cards, a box's tag and scores in its canvas table.
Procedures (slash commands: follow the skill, do not re-derive the steps)
| When the founder says | Run | What it does |
| "set me up" / a fresh copy | /get-started | the first run: questions about the business, config, segments and both canvases written, the example removed, the dashboard created, one first test proposed |
| "show me how this works" | /tour | six stops through a working loop: the example until set up, then the founder's own files |
| "test the idea that X" | /new-experiment | next number, folder, test card, canvas link, sync |
| "pull the numbers" / an experiment is Running | /pull-evidence NNNN | reads the sources the card names → a dated entry in the daily evidence log, compared with the bar |
| "log NNNN" / shares a result | /log-experiment NNNN | learning card, header done + decision, canvas box, one next hypothesis, sync |
| "sync the dashboard" | /sync-dashboard | builds the board, uploads the data files, creates the dashboard on the first run |
The Stop hook in .claude/settings.json runs dashboard/check_sync.py: when any
experiments/**/*.md is newer than the last sync, Claude is asked to sync before finishing, or to say
in one line why it is skipping.
Standing rules
- Nothing goes out without the founder's go, for that batch, in that chat. Building a list, a page or a draft is fine; sending, posting or publishing is their word.
- Personal data stays in CSV files only. Never an email address, phone number or name list in a
.mdfile: every.mdsyncs to the dashboard.experiments/**/*.csvis git-ignored on purpose. - Exclude test and internal accounts from every contact list and every count.
- Say when a window covers the weekend (
weekendinkit.config.json). - Never mark Persevere without a number. Interim reads are reads, not decisions; the learning card waits for the window to close.
- A Lean Canvas block is covered only with evidence: a row at Evidence 2 or more. A planned or running test does not count.
- One hypothesis at a time. After a Pivot or Kill, propose exactly one next hypothesis.
- Every test card plans its measurement before anything goes live: a "How we measure" table
naming, for each source, what it answers and the exact event, tag, filter or query. A card without
it is not ready.
sources:in the header lists the ones used.
Style rules
- One experiment = one small, falsifiable claim. A hypothesis with "and" in it gets split into two.
- Pick the cheapest experiment that still gives believable evidence. Prefer Discovery methods
(#1 Customer Interview, #12 Link Tracking, #13 Feature Stub, #14 404 Test) over Validation ones
(#35 Simple Landing Page is cheap but counts as Validation; #36 Crowdfunding, #38 Presale).
Write the method as
#NN Nameexactly as inMETHODS.md; if not in the 44,Custom (closest: #NN Name). - Priority order for new hypotheses: Desirability → Feasibility → Viability, unless the founder says otherwise.
- Replies to the founder stay short, plain and broken into small next steps.
Customer segments
customer-segments.md lists the working segments (Seg 1, Seg 2 …). Name the segment in every
hypothesis, in the choice of who to test on, and when reading a result. Do not assume a real customer
belongs to a segment unless the record says so.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

