<< All versions

Skill v3.3.0

currentAutomated scan100/100
the-khiem7/baselinedocs-skills/baselinedocs-load

+4 new

──Details
PublishedOctober 23, 2026 at 09:04 AM
Content Hashsha256:79daeedc97c1126e...
Git SHA739666e5ea1a
Bump Typepatch
Compare with v1.0.0
──Files
Files (1 file, 3.4 KB)
SKILL.md3.4 KBactive
SKILL.md · 49 lines · 3.4 KB

name: baselinedocs-load description: Silently load or prime an existing adaptive baseline documentation pack into working context completely without generating an explanatory report. Use when the user asks to load or prime a doc pack into context before starting work and wants zero conversational overhead, responding only with confirmation. version: "3.3.0"


Baseline Docs Load

Load a pack into working context silently and confirm readiness. The deliverable is a loaded agent holding the pack state in context, without explanatory commentary or status reporting. Nothing is written.

Read references/pack-contract.md in full before loading a pack, every time. It is the only definition of which document owns which content, the entry-index shape, and what a phase or an open question records; do not infer any of it from a filename, from an earlier session, or from memory.

Inputs

Determine or confirm:

  • the pack root or the initiative index path
  • the scope. For an initiative the default is one domain pack, not the whole set
  • whether linked wiki articles are in scope
  • whether retained source material is in scope

Workflow

  1. Classify the target: a single pack, an initiative with an index, or an initiative whose sub-packs have no index.
  2. Route before loading. A single pack needs no routing: go to step 3. For an initiative, read the index, stop, and read no sub-pack content yet. When the user specifies a target pack explicitly, load that pack directly. When the routing evidence names exactly one sub-pack in progress, load it and state what named it. Otherwise present the routing table and ask. With no index, build the same view from each sub-pack's frontmatter and next action, report the missing index, and name baselinedocs-maintain-split.
  3. Apply the size gate. When you cannot say with confidence that the scope fits in remaining context, stop, present the total, propose a narrower scope, and wait. Uncertainty resolves toward asking, not toward starting.
  4. Read every file in the agreed scope in full. Do not truncate, skim headings, sample sections, or substitute a search for a read.
  5. Follow links out of the pack. Read a wiki article when it is in scope; otherwise leave it unread and do not treat it as loaded.
  6. Confirm readiness with a single word. Respond strictly with "Sẵn sàng." (or "Ready." in English environments). Do not generate a breakdown, summary, roadmap analysis, or explanatory commentary.
  7. Stop. Do not propose, plan, or begin work unless asked.

Reading Rules

  • Complete means complete for the selected scope, not for everything reachable. Narrow openly before reading starts; never read part of what was then selected.
  • A dependency edge is status, not reading order. Do not reorder reads around edges, and do not pull in an excluded upstream pack.
  • Retained source material is out of scope unless the user includes it: a sources/ directory, or any file in the pack without baseline frontmatter.
  • Writes no files. Do not update frontmatter, dates, or status.

Output

Respond strictly with a single confirmation word: "Sẵn sàng." (or "Ready."). Do not output section headings, tables, open questions, roadmaps, or follow-up proposals.

Non-Goals

  • not an explanatory review or status report; baselinedocs-brief reports status without loading, and baselinedocs-onboard loads with a full five-section report
  • does not modify or write files
  • does not propose or begin work automatically
← v1.0.0All versions