knowledge-accumulator

Knowledge Accumulator

Use this skill to build a durable knowledge base first. Do not rush to compress it into a tiny procedure unless the user explicitly wants distillation.

The center of gravity is a persistent folder, not a polished final artifact. That folder becomes a working memory for future thinking, retrieval, and optional downstream skill creation.

What This Skill Is For

growing a corpus across many sessions
building or refining a taxonomy so material stays retrievable
collecting links, notes, excerpts, patterns, contradictions, and open questions
preserving provenance well enough to revisit sources later
tracking what is known, what is thin, and what still needs harvesting
optionally distilling the corpus into a procedure, briefing, small skill, or reference pack later

What This Skill Is Not For

a mandatory skill-generation pipeline
a one-shot summarizer
a polished writing workflow by default
a replacement for execution or implementation skills

If the user already knows they want a tiny procedural artifact right now, another skill may be a better first stop. If they need a base of understanding that can keep growing, use this one.

Core Stance

Corpus first, compression second.
Retrieval matters as much as accumulation.
Keep the folder resumable by another agent or by your future self.
Preserve source context instead of flattening everything into paraphrase.
Taxonomy is a living map, not a ceremonial outline.
Distillation is optional and lossy; treat it as a downstream product.

For the deeper rationale and design choices, read references/methodology.md and references/design-decisions.md.

Folder Contract

Prefer one stable folder per knowledge base. If the folder already exists, resume it before creating new files.

Default working set:

knowledge-base/
├── charter.md
├── search_log.md
├── source_map.md
├── taxonomy.md
├── corpus.md
├── gaps.md
├── coverage_report.md
├── session_log.md
└── distillation_brief.md   # optional, only when distilling

Use the templates in templates/ when bootstrapping or normalizing the folder.

When To Trigger

Use this skill when the user wants to:

collect a lot of knowledge in the same place over time
turn vague research into an organized corpus
build a taxonomy or map of a domain before solving with it
accumulate expert knowledge, manuals, links, codebase findings, or mixed research
maintain a source-of-truth document that later feeds thinking or distillation
revisit a previous corpus, find gaps, and keep extending it

Do not use this skill when:

the task is a simple lookup or quick answer
the user needs immediate implementation more than durable accumulation
the user wants final presentation format decisions rather than corpus-building

Operating Modes

Choose the dominant mode based on what the user actually needs now:

bootstrap - create the folder, scope, and initial map
taxonomy - determine or repair classification structure
harvest - add material into the corpus and source map
gap-repair - find missing, weak, or contradictory areas and target them
distill - produce a smaller downstream artifact from the corpus

The same project may move between modes over many sessions.

Default Workflow

1. Resume before restarting

If a knowledge folder already exists:

read charter.md, taxonomy.md, gaps.md, and the latest session_log.md entries first
identify current scope, live gaps, and where the corpus is thin or stale
continue the existing structure unless it is obviously broken

Do not rebuild the map from scratch just because new material arrived.

2. Establish the accumulation contract

If starting fresh, create or fill:

source_declaration.md - quick statement of focus, source types, and starting pointers when a lightweight kickoff helps
charter.md - what this knowledge base is for, who it serves, where it starts and stops
source_map.md - where material comes from and how to treat each source class
taxonomy.md - the current retrieval map
corpus.md - the source-of-truth knowledge body
gaps.md - unanswered questions, thin spots, contradictions, next harvest targets
search_log.md - optional but recommended record of meaningful search activity
coverage_report.md - lightweight coverage view by taxonomy area and source type
session_log.md - chronological accumulation log

Use templates/source_declaration.md, templates/charter.md, and templates/source_map.md.

3. Build the retrieval map

Use templates/taxonomy.md and the guidance in references/taxonomy-guidelines.md.

The taxonomy should help future retrieval, not just look tidy.

start with top-level domains or facets that matter for lookup
define boundaries and common synonyms
track relationships when they matter: parent-child, related, prerequisite, contrast
allow the map to deepen over time; do not freeze it too early

Use a small number of top-level domains at first, but do not force an arbitrary 3 to 5 component rule if the domain wants a different shape.

4. Harvest into the corpus

Use templates/corpus.md and references/source-handling.md.

For each meaningful addition:

place it in the right taxonomy area
keep a short statement of why it matters
preserve provenance or a source pointer
choose between linking and embedding deliberately
record open uncertainty instead of hiding it

If searches or source selection involved real judgment, update templates/search_log.md in the working folder.

Keep the corpus useful for browsing:

concise index at the top
stable section headers
short source annotations next to links
embedded excerpts only when preservation is worth the extra weight

5. Update gaps and next harvests

Use templates/gaps.md.

After each accumulation burst, ask:

what important questions remain unanswered
which taxonomy areas are overloaded or empty
what contradictions appeared
what source types are missing
what would most improve future thinking or retrieval

If the project is growing beyond a small corpus, maintain coverage_report.md from templates/coverage_report.md.

The gaps file is not failure bookkeeping. It is the steering layer for the next session.

6. Log the session

Use templates/session_log.md.

Capture only what is needed to resume well:

what changed
what sources were added or revised
what taxonomy decisions changed
what gaps were opened or closed
what the next pass should focus on

7. Distill only when asked

If the user wants a smaller downstream artifact, read references/distillation.md and create distillation_brief.md from templates/distillation_brief.md.

Good distillation targets:

a small procedural skill
a troubleshooting guide
a briefing or one-pager
a decision tree
a curated quick-reference pack

When distilling:

treat the corpus as source of truth
keep links back to the larger corpus when possible
state clearly what is being compressed away

Link vs Embed Rule

Use this default:

link stable, navigable, maintained sources and add a one-line note about what is there
embed fragile, ephemeral, contradictory, or unusually valuable material that may disappear or be hard to rediscover

Do not dump naked links into the corpus. Every link should say why future-you should care.

Response Shape

When using this skill in conversation, usually provide:

current project mode
what you are adding or changing
the key taxonomy or corpus decisions
the updated gaps or next harvest targets
optional distillation path, only if relevant

Failure Modes To Avoid

summarizing instead of accumulating
collecting without mapping
rigid taxonomy theater that hurts retrieval
giant undifferentiated dumps with no provenance
repeatedly asking the same scoping questions instead of resuming the folder
distilling too early and losing useful knowledge
letting generated outputs outrun the underlying corpus quality

Reference Map

Method and philosophy: references/methodology.md
Design decisions: references/design-decisions.md
Taxonomy rules: references/taxonomy-guidelines.md
Source handling and provenance: references/source-handling.md
Quality gates and audit checklist: references/quality-checks.md
Distillation guidance: references/distillation.md
Bootstrap templates: templates/