agentleFS
aalishan001aalishan001/opencockpit/.claude/skills/fourthwall/SKILL.md

fourthwall

Skill2 starsChanged 33 days agoMITView on GitHub

Documents the development and use of this cockpit itself (journal, changelog, README, case study) and produces a public, generalized, MIT-licensed copy of the repo. Commands - log, changelog, readme, export, publish, handoff. Use when the user talks about cockpit as a project, wants to record a decision or lesson about the system, or asks about the public version.

The file

Files
SKILL.md

FOURTHWALL

No command given (bare /fourthwall): read COMMANDS.md and print its ## FOURTHWALL section verbatim (heading through the line before the next ## ) in a fenced code block. Nothing else, no commit.

The domain that looks at the system itself: it keeps a private journal of how this cockpit was built and used, and generates a public, generalized copy that is safe for me and for anyone I work for. (This public repo was produced exactly this way.)

Principles

  1. Allowlist, never strip. The public copy is generated from an explicit manifest of generalized files (data/fourthwall/manifest.json). Nothing outside the manifest can reach it. Never copy or mirror the private repo, never make it public, never share git history: the public copy has its own fresh history.
  2. Hand off, don't write. Fourthwall never writes other domains' data. It uses their commands: bragdoc_stage.py add (the bragdoc dump path), ideas.py add, and a case-study draft queued for RUNWAY.
  3. Show then push, unless the owner set a standing rule. By default every publish shows me the full diff and the leak report and waits for my explicit approval. If I record a standing rule (state.json → auto_publish), a clean export whose own tests pass (python tests/test_suite.py in staging_dir) is approved and published without asking, and I'm told afterwards. Any failure means no publish.
  4. Excluded structurally: any private domain I list in the manifest rules (its skill, data, rules, scripts), data/, raw/, inbox/, out/, probes, profile.json, employment, real test fixtures, .git, settings with personal paths. The manifest rules reject them; validate.py and export enforce it.

Data

  • data/fourthwall/journal.jsonl: append-only {ts, kind: decision|lesson|milestone|bug, title, text, tags, commit?, source?, tier?, date_approx?}. Private source material; never exported.
  • data/fourthwall/state.json: last_changelog_commit, last_export_hash, approval, published_versions, ip_policy_checked (false until I confirm), commit_email, staging_dir, clone_dir, remote.
  • data/fourthwall/manifest.json: rules (excluded prefixes, substrings, names) + files: [{source, target, transform: copy|generalize|template, generalized?, source_hash?}].
    • copy: only for files with zero personal content (proven by the leak checks every export).
    • generalize: a hand-written rewrite in fourthwall/generalized/<target> (a target under .claude/ is stored under _claude/, so Claude Code never loads the copies as skills), tied to the source's hash. If the private source changes, export blocks as STALE: rewrite the generalized copy (remove my name, my personal assets, my employers, my data paths, private domains; keep the engineering intact), then python scripts/fourthwall.py mark-generalized <target>.
    • template: hand-maintained public files in fourthwall/templates/ (CLAUDE.md, LICENSE, CONTRIBUTING, .gitignore, data skeleton, fictional-persona examples, synthetic tests).
  • data/fourthwall/denylist.json: deny terms (see leak checks). Contact data is derived from profile.json at export time and never stored here.
  • fourthwall/CHANGELOG.md, fourthwall/README.md: generated, public-linted, shipped in the public copy.

Commands

python scripts/fourthwall.py does all exact work (ids, hashes, builds, checks, git). I write prose: journal entries, changelog summaries, README, case study, generalized copies.

CommandDoes
/fourthwall log <note>Infer kind and tags (ask only if unclear) → fourthwall.py log '<json>' → validate, commit fourthwall: log: <title>, push
/fourthwall changelogfourthwall.py changelog-commits → summarize the new commits into public-safe entries in my own words (never copy a commit message: they can name private items) → fourthwall.py changelog-append <version> '<json>' (rejects verbatim subjects, public-lint and deny-list failures) → commit, push
/fourthwall readmeRegenerate fourthwall/README.md from fourthwall/templates/README.template.md + journal + changelog: what it is, the harness-first idea with a plain-text diagram, the domains, the folder layout, how to add your own domain, safety rails, setup, how fourthwall builds the repo, license. No raw HTML and no mermaid: it must also render in a sanitized Markdown popup (a site's Repos section), whole if possible (about 6,000 characters of text). fourthwall.py lint-public fourthwall/README.md --voice must pass. Goes into the next export. Commit, push
/fourthwall exportfourthwall.py export builds the public copy into staging_dir (outside this repo, never committed) from the manifest only, then runs every leak check. Show me: file list, the full diff vs the last published version (fourthwall.py diff), the leak report. Never pushes. Stale generalized copies and leak failures block; fix them and export again
/fourthwall publishOnly after a clean export I approved in this session: fourthwall.py approve <hash> when I say so, then fourthwall.py publish <X.Y.Z> "<summary>" (copies staging into the local clone, commits release vX.Y.Z: <summary>, tags, pushes to remote). First publish is blocked until ip_policy_checked is true: ask me to confirm I've checked my employer's IP / side-project policy, then set-ip-checked. The clone commits with commit_email (a GitHub noreply address; never a personal email). After the first publish, tell me to turn on Settings → General → "Template repository". The script uses plain git only, never a hosting CLI that may be logged in as another account. Then python scripts/repos.py status: the new tag makes opencockpit's repo metadata due (EXTROVERT repos)
/fourthwall handoffQueue outward work without doing it: (1) python scripts/bragdoc_stage.py add '<json>': the public repo as a personal_project with links.repo, citing journal entries (I probe it later); (2) python scripts/ideas.py add -: the story of the repo, linking it (I run start later); (3) data/fourthwall/case_study.md (problem, constraints, architecture, decisions, results) for a RUNWAY showcase later, passing fourthwall.py lint-public data/fourthwall/case_study.md --voice. Validate, commit, push

Leak checks (export blocks on any failure; each reported pass/fail with offending lines)

  1. Structural: the staging tree holds exactly the manifest's targets; no excluded source; data/ targets only from templates.
  2. Secrets: detect-secrets scan of the staging tree.
  3. Public lint: disclosure_lint public channel on every text file (explicit per-file exemptions only in denylist.json → lint_exempt).
  4. Deny list (case-insensitive by default; short or common-word terms go in exact_terms, exact-case whole-word): my name variants, phone, emails, location (derived from profile.json), employers (employment.json), project names (staging), internal identifiers + excluded-employer terms (disclosure policy), clients, personal assets, user paths, private domains and their tools. My handle and name are allowed only on LICENSE and README author lines (author_line_exemptions).
  5. No em dashes anywhere in the public copy (prose uses colons, commas or parentheses; code writes the \u2014 escape).
  6. Canary (tests only): a planted fake secret, employer name and private-domain name in a temp copy must all be caught.

The journal is private, but anything drawn from it (README, changelog, case study) passes the public lint and the deny list.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.