vault-bootstrap
Skill2 starsChanged 55 days agoMITView on GitHub
Initialise an AI-directed Vault in a repository from VAULT-STARTER.md, or add a new axis to a repo that already has one. Use when the user says "set up a vault here", "initialise the vault", "read VAULT-STARTER and bootstrap this project", "add a second axis", or when a project has months of context living in chat and no durable place to put it. Covers the interview, the directory layout, the seed files, the privacy wiring and the baseline audit. Do NOT use it to reorganise a vault that already exists (that is vault-lint) or to close a session (that is vault-checkpoint).
In this repository2 other files it gives its agents.
- vault-checkpointSkill
- vault-lintSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
14 rules. About 0 are common, 8 are distinctive, and 6 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
6
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 23
Read VAULT-STARTER.md before starting.
- Distinctive: line 28
Ask all of it in a single round.
- Distinctive: line 41
If several: which is the default when a session is ambiguous, and which ones are private.
- Distinctive: line 49
Follow §3. Respect whatever already exists.
- Very specific: line 79
Lecciones.md: start with the Environment section (build, test, paths). - Distinctive: line 89
Run it once for a baseline.
- Distinctive: line 93
If there is a git repo, add glob patterns, never literal lines:
- Very specific: line 103
Never put credentials in any .md, only references to where they live.
- Very specific: line 107
If sources already exist, run one Ingest Loop to compile them into 10-Knowledge/.
- Distinctive: line 112
Use the block in §11. It must name the axes, the default one, and the language boundary.
- Distinctive: line 117
Run the closing routine (§7): commit, and show the auditor's report as the startup-weight baseline.
- Distinctive: line 122
Do not build what §8 defers: no scheduler, no declarative multi-agent orchestration, no embeddings.
- Very specific: line 125
Do not auto-load knowledge pages via
@import: Only what every single session needs. - Very specific: line 127
Do not invent project state: If you do not know where the project is, ask.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill 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
Bootstrap a Vault
What this does
Turns a repository into one the agent can work from for months without
re-explaining the project every session. The deliverable is a directory of
markdown plus two seed habits: every session starts at Current-State.md, and
every operation ends by updating three files.
Read VAULT-STARTER.md before starting. This skill is the procedure; that
document is the contract.
Step 1 — Interview, once
Ask all of it in a single round. A bootstrap that interrupts six times gets abandoned.
- Project name and domain. Software, research, writing, operations, anything. The domain decides the loop names.
- Existing raw sources. README, specs, transcripts, design docs. Anything that already holds truth.
- How a deliverable is validated here. Tests, review, acceptance criteria. Then the question most people have not thought about: what control case proves that validation actually measures anything? (VAULT-STARTER §8.5.)
- Individual or collaborative. This changes the optimisation strategy from day one (§9.4).
- One axis or several. A repo can hold several independent vaults (§3.6). If several: which is the default when a session is ambiguous, and which ones are private.
- Language boundary. What is written in the working language and what in the audience's language (§5.6). Deciding late leaves translation debt nobody has sized.
Step 2 — Create the structure
Follow §3. Respect whatever already exists; never move a user's files to fit the template.
<repo>/
├── CLAUDE.md
├── VAULT-STARTER.md
└── Vault/ # or <AxisName>/ per axis in a multi-axis repo
├── SCHEMA.md
├── 00-Index.md
├── LOG.md
├── 10-Knowledge/
├── 20-State/
│ ├── Current-State.md
│ ├── Task-Board.md
│ ├── Lecciones.md
│ └── Decisiones/
├── 30-Loops/
└── 90-Raw/Step 3 — Seed the files
SCHEMA.md: sections 2 to 9 ofVAULT-STARTER.md, with loop names adapted to the domain.Current-State.md: the real current state, plus a closing section "NEXT SESSION" listing what comes next and which decisions are waiting on the human. Ceiling 2,500 tokens.LOG.md: one entry,state | Vault initialised.00-Index.md: one line per page, grouped by layer.Lecciones.md: start with the Environment section (build, test, paths).30-Loops/: the five base loops from §6.
Step 4 — Install the auditor
Copy scripts/check_vault.py without editing it. It discovers axes by
filename, so it works unmodified. If a ceiling genuinely needs to move, create
vault-config.json rather than editing the script (§7.1) — an edited copy is
how the ceiling drifts and nobody notices.
Run it once for a baseline.
Step 5 — Wire privacy
If there is a git repo, add glob patterns, never literal lines:
Notas-Privadas*
Bitacora-Privada*Plus any fully private axis directory. Verify with git check-ignore -q, not
by reading .gitignore — a glob does not appear as a literal string.
Never put credentials in any .md, only references to where they live.
Step 6 — First ingest
If sources already exist, run one Ingest Loop to compile them into
10-Knowledge/. Flag contradictions explicitly; never resolve one silently.
Step 7 — Write CLAUDE.md
Use the block in §11. It must name the axes, the default one, and the language boundary.
Step 8 — Close
Run the closing routine (§7): commit, and show the auditor's report as the startup-weight baseline.
Anti-goals
- Do not build what §8 defers: no scheduler, no declarative multi-agent orchestration, no embeddings. Under roughly 300 pages the index replaces search.
- Do not auto-load knowledge pages via
@import. Only what every single session needs. Everything else lives one path away. - Do not invent project state. If you do not know where the project is, ask.
A confident but wrong
Current-State.mdis worse than an empty one.
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.

