opencockpit
CLAUDE.md2 starsChanged 33 days agoMITView on GitHub
Personal command system run inside Claude Code. Five domains, each a skill with explicit commands. "Agent" here means a division of duties, not a separate process. /end ends whichever interactive command is active (conversation, editorial, probe) and commits. If a request doesn't name a command, map it to the closest one and say which you used. If it maps to nothing, say so and suggest the nearest command. Don't improvise a new workflow. 1. No invented facts. Numbers, dates, metrics,…
- Reads credentials
- Commits and pushes
The file
We split this CLAUDE.md into the rules it gives an agent
7 rules. About 0 are common, 2 are distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 17
If a request doesn't name a command, map it to the closest one and say which you used.
- Distinctive: line 18
If it maps to nothing, say so and suggest the nearest command.
- Very specific: line 35
Provenance: Every record in data/ carries a source (file name in raw/, web:<url>, or user).
- Very specific: line 38
Show then write: Before changing data/, show what will change (a short diff or table).
- Very specific: line 40
Versioning = git + snapshots: After every command that changes files, commit: <domain>: <command>: <short summary>.
- Very specific: line 55
Identity: before the first push in a session, confirm the repo's local user.email and origin URL are the ones I set for this repo.
- Very specific: line 60
Keep this repo private: data/ and raw/ hold your real records.
≈ “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
cockpit
Personal command system run inside Claude Code. Five domains, each a skill with explicit commands. "Agent" here means a division of duties, not a separate process.
| Domain | Skill | Purpose |
| CHASSIS | /chassis | Example domain: vehicle maintenance from the owner's manual, service slips and odometer readings |
| EXTROVERT | /extrovert | Interview-style conversations → edited posts (dev.to canonical, published through its API; LinkedIn summary), ideas backlog, profile copy |
| bragDoc | /bragdoc | Probed, metric-backed record of engineering work → bragDoc + extensive resume, with disclosure tiers |
| RUNWAY | /runway | Ships the bragDoc's public channel to your own site repo (preview on a content branch, promote on your approval) |
| FOURTHWALL | /fourthwall | Documents the system itself (journal, changelog, README, case study) and exports a public, generalized copy |
| HELP | /cockpit [domain] | Prints COMMANDS.md (or one domain's section) verbatim |
/end ends whichever interactive command is active (conversation, editorial, probe) and commits.
If a request doesn't name a command, map it to the closest one and say which you used. If it maps to nothing, say so and suggest the nearest command. Don't improvise a new workflow.
Folder contract
inbox/: drop zone (create it). Anything here is unprocessed.- Optional second inbox (a synced cloud folder for phone uploads): set
COCKPIT_DRIVE_INBOX. Same rules asinbox/; check both.python scripts/ingest_images.py --pendinglists images waiting in either. raw/: originals after processing, moved from inbox asraw/<domain>/<YYYYMMDD-HHMM>-<name>. Never delete.- Images are always ingested via
ingest_images.py, never committed as raw phone photos.python scripts/ingest_images.py <domain> <paths>writes compressed JPEGs toraw/<domain>/and deletes the inbox originals (the raw/ copy is the archive). data/: structured truth. Only skills write here, only in the formats inscripts/schemas.md.templates/: rendering templates for PDFs and HTML.out/: rendered outputs (PDF/HTML). Regenerable; always produced by scripts.examples/persona/: a fictional user's data. Copy it intodata/to try every command before using your own.- Your website repo (optional): a sibling of this repo, never inside it, set in
data/runway/config.json. Only RUNWAY writes there, one way: content (the site data and the resume PDF) through publish, and site code changes you ask for through ship.
Global rules (all domains)
- No invented facts. Numbers, dates, metrics, odometer readings, and claims come from me, an image or document I gave, or a cited web source. If unsure, ask. Never fill a gap with a plausible guess and store it as fact.
- Provenance. Every record in
data/carries asource(file name inraw/,web:<url>, oruser). - Ask, don't guess. Low-confidence extractions go to the question queue for that domain and are asked at the end of the command. They never go into
data/until answered. - Code does exact work. Arithmetic, dates, sorting, deadlines, hashes, rendering, and validation are done by
scripts/. Never compute these in your head. Run the script and report its output. - Show then write. Before changing
data/, show what will change (a short diff or table). For bulk changes, show a summary and ask once. - Validation is enforced. A hook runs
scripts/validate.pyafter every edit. If it fails, fix the data before doing anything else. - Versioning = git + snapshots. After every command that changes files, commit:
<domain>: <command>: <short summary>. Conversations, editorials and probe transcripts are also saved as dated files (each skill names its paths), referenceable by date-time and by name. - Commands end the way their spec says. One-off commands finish after output. Interactive commands continue until I say
/endor their auto-end condition is met. - Tone. Direct, concise, no flattery. If my claim or plan is weak, say so.
- Command names. COMMANDS.md is the source of truth for command names. Whenever a skill's commands change, update COMMANDS.md in the same commit.
- Public data. Only RUNWAY publishes, and nothing goes live until I approve a preview (
/runway promote). Visibility of a bragDoc entry changes only on my explicit instruction: a specific one (bragdoc_public.py set-visibility), or my standing default indata/runway/config.json(applied once, when renovate creates an entry). Public summaries and site cards are generated automatically and must follow the redaction rules in the bragdoc skill. - Disclosure channels. Anything that leaves the bragDoc is internal, resume or public, as defined in
data/bragdoc/disclosure_policy.json. Every resume and public output passesscripts/disclosure_lint.pyfor its channel before I see it; nothing ships with a lint failure. - Names. Read them from
data/profile.json→names; never hard-code one. Internal output:short. Public channel (website, posts, LICENSE, README, case studies, social):shortby default,public_fullonly where a full name is genuinely required, such as a copyright line or author credit. Resume channel (resumes and formal applications only):formal, exactly as stored. An all-caps design is CSStext-transform: uppercase; the stored and linted text never changes.
Git operations
I never touch git. You own it, under these rules:
- Branch:
mainonly. No branches, no PRs. - Session start:
git pull --rebase(the SessionStart hook does this; if it failed, do it and resolve conflicts, preferring the newer data line and showing me). - After every command that changes files: validate, then commit
<domain>: <command>: <summary>andgit pushimmediately. No batching across commands, no asking. - Identity: before the first push in a session, confirm the repo's local
user.emailandoriginURL are the ones I set for this repo. If not, stop and tell me. Never fix it by changing global config or credentials. - Never: force push, rewrite pushed history,
reset --hard, delete branches (one exception: RUNWAY deletes a site branch once it is fully merged intomain, rule 7), touch global git config, or commit secrets/tokens/.env. - Push failures (auth, network, rejected): report the exact error in one line and keep working locally; retry push at the next commit.
- Website repo (if you use RUNWAY): the same rules, through
scripts/runway.pyonly. Content:publish→ branchcontent(preview) →promote. Site code (only on my explicit instruction):ship "<summary>"→ its own branchcode/<slug>(preview) →promote code/<slug>. Both paths run the same gates, and nothing reachesmain(live) without my approval of the preview. A preview branch exists only while something waits for my approval: every promote deletes the branch it promoted, andrunway.py prunedeletes any other branch already merged intomain; a branch with commits not inmainis never deleted without showing me. The deploy hook lives only in the gitignoreddata/runway/deploy-hook.local.
Keep this repo private: data/ and raw/ hold your real records. Share only what FOURTHWALL exports.
Profile
See data/profile.json. Tenure baseline for "how impressive is this" judgments: career start in that file → today.
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.

