Overview
The recurring vault tasks are stable enough to be fixed prompts rather than freehand asks. Each one below has a defined input, a defined output shape, and a rule about what it may not do. They are the working parts of the loop in vault-orchestration. Keep them in the vault, under 90-Templates/ or 99-Tools/, and version them with it.
Keep the prompts fixed so runs are comparable
A freehand “extract the key ideas from this chapter” produces a different shape every time, so nothing can be compared across sessions and nothing can be audited. A fixed prompt with a fixed output contract makes the second run measurable against the first, which is the same reason the judging rubric in semantic-audit is fixed. Edit a prompt through the same review path as any other vault change, and note the date it changed.
Extract atoms under a contract, not a request
Input: the source text, plus the list of existing atom titles the extraction may link to. Output, per atom:
- A title written as a full claim sentence, kebab-cased for the filename.
- A body of two to five sentences, paraphrased in the owner’s voice, never copied.
- The exact supporting quote with its location: page, timestamp, or section.
- A confidence rating with a one-line rationale.
- Two or three candidate links drawn only from the supplied title list.
Emit each atom as a separate block ready to paste into 20-Atoms/. The rules the block must satisfy are the ones in atomic-notes: one idea per note, and a title that needs “and” is two atoms.
Let maps emerge from clusters, not from a plan
Input: N atom titles from one domain. Output: clusters of four or more, a proposed map title for each, and a one-sentence rationale per member. Clusters smaller than four are not yet a map. Growing maps from clusters rather than declaring them up front is the policy in linking-and-tags.
Ask for the seam, not a verdict
Input: one note. Output: either the sentence where the note pivots from its first claim to its second, or a statement that it holds one idea. Asking for the seam forces evidence; asking “is this atomic?” invites a guess.
Require a relationship type on every proposed link
Input: a new atom and up to fifty candidate titles. Output: three to seven outbound links, ranked, each with a one-sentence rationale and a relationship type: supports, refines, contradicts, generalizes, specializes, depends on, or parallel to.
If fewer than three candidates genuinely connect, the prompt says so. Padding a link list to hit a count creates the naked link dump the auditor flags.
Detect contradictions as pairs with a draft tension note
Input: a domain’s atoms. Output: pairs that cannot both be true, and for each pair a draft contradiction note naming the atoms in tension and a resolution criterion. The prompt proposes the tension; it never resolves it.
Triage the audit report into batch actions
Input: the auditor’s JSON report. Output: findings grouped by likely root cause, with one batch action per cluster and criticals listed individually first. This is the prompt that turns thirty findings into four sessions, per vault-audit.
Run a pre-mortem before a high-stakes decision
Input: the draft decision entry. Output: assume it is eighteen months later and the decision failed, then the three likeliest causes, with the checkable ones rewritten as assumptions to add to key_assumptions. The decision entry itself is described in decision-journals.
Chain the prompts; do not merge them
Extraction feeds link suggestion, which feeds the atomicity check, which feeds the audit. Each stage’s output is the next stage’s input and each is reviewable on its own. Merging them into one large prompt removes the handoff points that keep the agent from writing unverified claims straight into the vault. The chaining pattern is the standard one in prompt-chaining.