Creating Distillations

Creating Distillations

Per-distillation projection of one source onto one task axis. Consumes a (source, task) pair plus an optional lens-set; produces one file plus an index update.

Invoked by:

• creating-applications, per-distillation orchestration: once per applicable (source, task) distillation.
• ingesting-resources Pass G: once per task axis during a new-source ingest.

Rationale, template structure, evidence discipline, four-outcome lens-applicability gate, and failure modes live in projection-protocol.md. This file is the runbook; that one is the rationale. Read both before invoking.

When to use

• Pass G of ingesting-resources. Per task axis in the corpus, the ingest skill calls this skill once with the new (source, task) pair. The applicability gate decides whether to produce a distillation or skip.
• Per-distillation orchestration in creating-applications. When a new application is being assembled, the orchestrator calls this skill once per (source, task) distillation named in the task spec. Often dispatched as parallel sub-agents.
• Operator follow-up. A new task axis has landed and the operator notices a source whose existing deep reference predates the axis: invoke this skill directly to project that source onto the new task.
• Re-projection after a deep-ref update. The deep reference has been sharpened (often via Pass G cycling: see projection-protocol.md §Cycling); the distillation needs to reflect the updated deep.

When not to use

• The deep reference does not yet exist. Run ingesting-resources first.
• The task axis does not yet exist (no distillation directory, no index, no build profile). Run creating-tasks → creating-applications first.
• The operator wants to project all applicable sources onto a new axis. That's the per-distillation orchestration job inside creating-applications: invoke it instead, and it will call this skill once per source.

Pre-flight

Both checks must pass:

1. Verified deep reference exists. Path: corpus.commons/{corpus}/references/{slug}-deep.md. The deep must have passed Pass I (source-only audit). If the deep does not exist, abort with the operator-facing message "Deep reference for {slug} not found; run ingesting-resources first."
2. Task axis exists. Path: corpus.commons/{corpus}/distillations/{task-name}/. The directory plus task-index.json (and the human-readable {TASK-UPPERCASE}-DISTILLATION-INDEX.md operator-inspection view) must exist. If any are missing, abort with "Task axis {task-name} not yet scaffolded; run creating-applications first."

If both pre-flight checks pass, the skill proceeds. Otherwise it fails loudly and names the gate that closed.

Inputs

• The existing task axis name in kebab-case (e.g., decision-making, learning-from-incidents).
• The slug of the existing verified reference (e.g., openstax-organizational-behavior).
• Optionally, one or more lens names. Lens names refer to specs at corpus.commons/{corpus}/lenses/{lens-slug}.md. If a named lens has no spec file, abort and direct the operator to run creating-lenses first.

Outputs

• The distillation file at corpus.commons/{corpus}/distillations/{task-name}/{slug}-{task-name}.md, produced when Pass G.0's source-applicability gate fires clear yes or operator-confirmed yes. The distillation includes a ## Runtime triggers this source addresses section that maps the source's content to specific triggers from the task spec's seed table (field 2a), with teach-in-the-moment scripts. See "Trigger-grain projection" below.
• An update to the distillation index for the task, citing the new distillation in the appropriate phase sections and adding rows to the per-phase runtime listener tables for each trigger the source addresses.
• For each lens where the per-distillation applicability gate fires beyond clear no:
  • Partial reshape: a ## Through the {lens-name} lens section inside the main distillation.
  • Deep reshape: a separate per-lens distillation at corpus.commons/{corpus}/distillations/{task-name}/{slug}-{task-name}-{lens-name}.md.
• A skip log when Pass G.0 fires clear no or operator-confirmed no. No file is generated; the skip is recorded with the reason. Lens skips at clear no are recorded similarly.

If the skill is invoked under parallel-batch orchestration, index updates may be written to a staging file (_ingest_{task}_{slug}.md in the distillation directory) instead of the canonical index, to be merged by the orchestrator's consolidator step.

Procedure

The detailed step-by-step procedure—applicability gate logic, projection-template structure, evidence discipline, lens-as-optional-framing—lives in projection-protocol.md. The runbook summary:

1. Read the verified deep reference. The deep is the only input at this step; the source is not re-read. (See projection-protocol.md §Failure modes for why.)
2. Run the source-applicability gate (projection-protocol.md §Pass G.0). Three outcomes: clear yes → proceed; clear no → log skip and exit; ambiguous → ask the operator and proceed per their call.
3. Project the source onto the task using the standard nine-section template (projection-protocol.md §Projection template). The structure is uniform across the corpus; do not deviate, add sections, or reorder.
4. For each named lens, run the lens-applicability gate (projection-protocol.md §Lens-as-optional-framing). Four outcomes:
   • Clear no. No additional artefact. Log the skip.
   • Clear yes, partial reshape. Add a ## Through the {lens-name} lens section inside the distillation.
   • Clear yes, deep reshape. Produce a separate per-lens distillation file.
   • Ambiguous. Ask the operator; proceed per their call.
5. Trigger-grain projection. Read the task spec's field 2a (seed trigger→response table) and identify which triggers this source can address. For each:
   • Add a row to the distillation's ## Runtime triggers this source addresses section: trigger | source's content that addresses it | teach-in-the-moment script (1-3 sentences).
   • Add a row to the distillation index's runtime listener table for the appropriate phase: trigger | framework name + cite to this distillation.
   • The source may extend the seed table (surface triggers the spec didn't anticipate but the source clearly addresses); log these extensions in a ## Trigger extensions surfaced by this source section so the operator can fold them back into the task spec.
   • If the source addresses zero triggers in the seed table, that is a signal Pass G.0 should reconsider: the source may apply at the phase level but not at the moment level, which means the distillation will be shelf material at runtime even if the phase-routing fires it. Default: still produce the distillation but flag the trigger-mismatch in the distillation header so the operator can address it.
6. Update the distillation index to cite the new file(s) in the phase sections where they apply and add trigger-table rows for the triggers the source addresses. If running under parallel-batch orchestration, write to a staging file instead and let the orchestrator's consolidator merge it.
7. Run npm run build to verify the affected profile compiles cleanly. Required after every Mode-2 run when invoked standalone; the orchestrator runs a final build after all per-distillation invocations complete when running under creating-applications.

Re-runs and overwrites

If the distillation already exists for the (source, task) pair, the skill does not silently overwrite. Two paths:

• Refresh after a deep-ref update. Delete or rename the existing distillation; rerun the skill. The applicability check runs again; the file is regenerated from the current deep.
• Merge changes. The skill does not merge. Generate the new version into a temporary path, diff against the existing, and apply the diff manually.

Failure modes to watch

The full catalogue with worked examples lives in projection-protocol.md §Failure modes. The headlines:

• Skipping the source-applicability gate. Generating a distillation for a (source, task) pair where the source doesn't actually fit produces a file that strains, reaches, and reads as filler. Better to skip and record the skip.
• Skipping the lens-applicability gate. Pre-projecting against every named lens by default inflates the corpus with thin lens-shaped variations of the same content. The default for lenses is do nothing; pre-project only where the lens materially reweights the distillation.
• Re-reading the source. Distillation projection is from the verified deep, not from the source. Re-reading reintroduces failure modes the deep ref has already audited out (training-prior leakage, drift, evidence-marker confusion).
• Smuggling new claims into distillations. If a concept is not in the deep reference, it does not belong in the distillation. Re-ingest if the concept warrants inclusion.
• Re-extracting verbatim from the source at Pass G. Verbatim blockquotes in distillations are copies of already-audited Pass D blockquotes from the deep ref, with the evidence marker preserved. Re-extracting from the source reopens the failure modes Pass D closed.
• Mismatched task vocabulary. The distillation's working vocabulary should be the task's own: decisions for decision-making, stakeholders for stakeholder-engagement, incidents for incident-review, etc. Source-author vocabulary is preserved in the deep reference; task-applied vocabulary lives in the distillations.

Related skills

• creating-tasks: scopes a new task axis. Run before this skill if the named task axis does not yet exist.
• creating-applications: assembles a new application. Invokes this skill once per (source, task) distillation.
• ingesting-resources: runs the 9-pass protocol. Pass G calls this skill once per task axis the corpus carries.
• creating-lenses: designs a lens spec. Run before this skill if a named lens doesn't yet exist.
• matching-references: find an existing source by topic before assuming a distillation is needed.