What does this skill do, and when should you use it?
Factory Floor is an open-source Claude Code skill by Swiftner for founders and early-stage teams. It applies the procedures and thresholds of Theory of Constraints (Goldratt), the Customer Factory (Maurya), and Jobs To Be Done (Christensen/Moesta/Ulwick) to prioritization and bottleneck diagnosis. Every request is sorted into two modes: 'Procedure' questions with checkable answers (funnel arithmetic, WIP limits, weekly reviews) get maximally directive answers, while 'Judgment' questions (what's the constraint, should we pivot) never get prescriptions — instead the skill takes your theory first and designs a discriminating test. A state layer in .factory/context.md and journal.md persists your stage, numbers, and experiments across sessions, with a prediction ledger that scores your forecasts each quarter. The whole skill is ten files, about 60KB, with no scripts or external dependencies.
Routes to one of four stage playbooks (pre-revenue / restart / growth / scaling) via triage; runs named procedures including funnel break scans, weekly reviews, napkin math, WIP audits, interview scripts, and positioning sprints; keeps all feedback anchored on venture numbers rather than the founder; extracts a prediction with a confidence number before commenting and writes it to .factory/journal.md, scoring the outcome at the experiment's deadline; maintains .factory/context.md with a fixed H2 schema where the current constraint must name one factory step — Acquisition, Activation, Revenue, Retention, or Referral — plus magnitude and date.
- A pre-revenue founder torn between writing code and finding customers who needs validation procedures, not a feature list
- A small team that once had customers and lost them all, wanting churned-customer interviews and the restart sequence
- A company under $100K MRR with fewer than 10 people that wants to find the constraint and run two-week experiment cycles
- A team juggling five parallel projects that ships nothing and needs a WIP audit forcing convergence on one experiment
- A founder talking up hiring or fundraising despite flat growth — the skill asks about churn and net revenue retention first
- A Claude Desktop user who pastes SKILL.md as project Custom Instructions and uploads the ten reference files as Project Knowledge
How do you install this skill?
- Static review only, nothing executed; references/, stages/, and scripts/ files are absent from the evidence, so actual routing and state-writing behavior are unverified.
- Version mismatch between _meta. (3.4.1) and package. (4.1.0); verify the actual published version before installing.
- The skill writes .factory/ state files into your working directory with no stated rollback path; be aware in shared repositories.
- No Chinese-language support declared; docs are English-only. No overseas-service dependency, so mainland-China reachability is not a concern.
- Cited research figures (e.g., +34%/−8%) are author paraphrases without primary sources; verify independently.
- Local filesystem
Claude Code / OpenAI Codex: run npx @swiftner/factory-floor in your terminal — it installs to ~/.claude/skills/factory-floor/ or ~/.codex/skills/factory-floor/ respectively. Alternatively via ClawHub: clawhub install factory-floor. Claude Desktop (manual): create a new Project, paste SKILL.md as Custom Instructions, and upload the four stage files plus the five reference files (intake, probes, playbooks, weekly-review, state) as Project Knowledge. No license is stated in the repo, so verify before adopting.
git clone --depth 1 https://github.com/Swiftner/Factory-Floor.git ~/.claude/skills/Factory-FloorGenerated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Should we build Slack integration?
- What should we work on this week?
In Claude Code, just talk about priorities, bottlenecks, or flat growth — e.g. 'Should we build Slack integration?' or 'What should we work on this week?' — and it triggers automatically. It first triages your stage and reads .factory/context.md and journal.md (first run goes through intake questions). In Claude Desktop, paste your context.md at the start of each session and copy suggested state updates back into your project files. After triggering, it asks for your prediction first, then either runs the named procedure or designs a discriminating test, and assigns actions with concrete day-and-time deadlines.
What are this skill's strengths and limitations?
- Grounded in coherent frameworks (Theory of Constraints, JTBD, Customer Factory) with concrete numeric thresholds instead of vague advice
- The two-mode design is evidence-based: assistants lift novices most on structured, checkable work and hurt them on ambiguous judgment calls
- The prediction ledger forces commitment before action and mandatory scoring at deadlines, countering founder overconfidence
- A filesystem state layer means you never re-explain yourself across sessions
- Tiny footprint — ten files, ~60KB, no scripts or external dependencies — easy to audit and port
- No license is declared in the repository, creating legal uncertainty for commercial use or modification
- Claude Desktop lacks a filesystem, so the state layer must be maintained by manual copy-paste
- Strictly startup-scoped; the README explicitly excludes freelancers, agencies, and established companies
- The quality of the stage and reference files can't be verified from the README, and there is no test suite
- The style is deliberately blunt (no preamble, never validates plans you're attached to), which some users will find abrasive
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Factory Floor — Startup Coach this page | 53 · Use with care | ★ 14 | 2mo ago | — |
| vibe-check Product Planning Coach | 62 · Recommended | ★ 619 | 3mo ago | MIT |
| Grow Sustainably Advisor | 49 · Use with care | ★ 11k | 6mo ago | — |
| Interview Coach | 50 · Use with care | ★ 2.4k | 4mo ago | MIT |
| Product Manager Skills | 59 · Recommended | ★ 186 | 6mo ago | NOASSERTION |
How did FollowSkills review this skill?
Purely advisory skill: no network calls, no credential access, state written only to local .factory/ files with transparent data flow; but no stated write-confirmation or rollback mechanism, and SKILL.md itself carries no license (MIT appears only in package.). 9 points deducted.
Routing logic, state schema, and abnormal-input handling (missing/renamed headers) are clearly specified with good failure feedback; but this is static review — the referenced references/ and stages/ files are absent from evidence, so key paths cannot be reproduced. 11 points deducted.
Trigger phrases, audience, and explicit non-fit ranges (NOT for list) are concrete; stage thresholds are semantically clear; but English-only with no Chinese support declared, and broad trigger phrases risk mild over-triggering. 4 points deducted.
Layered docs, progressive disclosure, content-ownership table, and maintenance principles are strong; version and repo declared; but _meta. (3.4.1) conflicts with package. (4.1.0), no CHANGELOG, and maintenance responsibility rests on a signature only. 6 points deducted.
Goals, sample outputs, and design rationale (two-mode sort, prediction ledger) are well stated with named sources; but static review cannot verify live session quality, no representative outputs are evidenced as directly usable, and marginal value rests on user trust. 8 points deducted.
Framework credits are detailed and README gives concrete examples; but core claims (routing behavior, state machine, +34%/−8% experiment figures) lack tests or third-party execution evidence, and fact/inference boundaries are blurred. 7 points deducted.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →