| name | progressive-disclosure |
| description | Progressive-disclosure + install logic for generated context packs: the SKILL.md token-split threshold and measurement, the uniform index schema, and the write+copy-back install scripts. Keywords: progressive disclosure, split threshold, SKILL.md index, references, install script, copy-back, round-trip. |
Progressive Disclosure
How a generated context pack stays a lean entry point: the token-split
threshold, the uniform index schema once split, and the write+copy-back install
scripts. Loaded by the indexer (and the writer for the index schema).
When to Use This Skill
Load this skill when:
- Measuring a
SKILL.md body and deciding whether to split (indexer).
- Restructuring a
SKILL.md into an index + references/01..05 (indexer).
- Generating the
_install/ copy-back scripts (indexer).
Quick Start
- Measure the
SKILL.md body (frontmatter excluded). See
references/split-threshold.md — the
single source of truth for the threshold constant (5,000 tokens) and
the deterministic heuristic.
- If over threshold → split into the uniform index +
references/01..05. See
references/skill-index-schema.md.
- Generate
_install/install-context-pack.{sh,ps1}. See
references/install-roundtrip.md — the
single source so the two scripts cannot drift.
Core Concepts
Progressive disclosure
A SKILL.md should be a lean entry point that loads detail on demand from
bundled references/*.md. Keeping the index small saves context tokens when the
harness auto-loads the pack. The split is re-evaluated every run: a pack can
cross the threshold as it grows, or uncross it as it shrinks (then references/
collapse back into SKILL.md).
Split threshold (centralised)
The threshold is 5,000 tokens of body. It lives in one place —
references/split-threshold.md — so a correction
(e.g. after re-verifying live docs) is a one-line change. The conformance eval
asserts no shipped index SKILL.md body exceeds it.
Round-trip install
The pack is written to the context repo; the generated _install/ scripts
copy it back into a code repo's harness skills dir when the user runs
them. No agent ever runs the scripts.
Best Practices
- Centralise the number — never hardcode the threshold in the agent prompt;
read it from
split-threshold.md.
- Index stays under threshold — the whole point of splitting.
- Two native scripts, one logic source —
.sh + .ps1 share identical
behaviour described in install-roundtrip.md.
- Agent generates, user runs — the indexer never executes the scripts.
Quality Checklist
References