Knowledge Curator

Knowledge Curator

Not every useful insight belongs in a one-off chat. Knowledge Curator turns repeated lessons from session history into maintainable project guidance so future sessions start from better defaults.

Why This is Copilot-Exclusive

Copilot CLI gives you three building blocks that work well together:

• Session artifacts for saving structured notes during a task
• /resume for continuing a past session with its artifacts and SQL state
• session_store search for finding prior discussions and patterns later

The key constraint matters: session_store is read-only historical memory, not a writable knowledge base. This skill is about deciding what should graduate from temporary session context into durable project docs such as .github/copilot-instructions.md, CONTRIBUTING.md, or reference guides.

When to Use

• The same repo-specific warning or pattern keeps appearing across sessions
• A temporary workaround has become a stable team practice
• You want to prune stale guidance instead of accumulating contradictory notes
• A sprint or migration produced lessons worth promoting into project instructions

When NOT to Use

Prerequisites

• A candidate insight worth keeping
• A target destination for that insight
• Enough evidence to tell whether it is stable, repeated, or obsolete

Knowledge Layers

Workflow

1. Capture candidate lessons during the task

When you notice a reusable pattern or pitfall, save it in the current session workspace instead of burying it in scrollback:

Save a short session note with:
- the pattern or pitfall
- where we saw it
- whether it is provisional or stable

Keep these notes short and concrete.

2. Review historical evidence

Search prior sessions before promoting anything:

Search prior sessions for notes about flaky setup, markdown lint failures, or validation order.
Summarize the repeated findings before we decide what to keep.

Promotion is stronger when the same lesson shows up more than once.

3. Promote only durable guidance

Move proven guidance into the right permanent home:

• .github/copilot-instructions.md for broad repo-wide defaults
• CONTRIBUTING.md for contributor workflow expectations
• guides/ or references/ for richer explanations and examples

Example prompt:

Promote the validated markdown workflow lessons into CONTRIBUTING.md.
Keep the wording short, actionable, and repository-specific.

4. Prune or rewrite stale guidance

Do not only add. Review for drift:

Compare the current instructions against what recent sessions actually required.
List any stale, duplicated, or contradictory guidance before editing.

If a note is no longer true, remove or rewrite it instead of layering on exceptions.

5. Export a compact handoff when needed

At the end of a sprint or migration, create a concise handoff artifact:

Create a short project-knowledge summary for this sprint:
- 5 lessons to keep
- 3 pitfalls to avoid next time
- 3 items that should remain temporary

Use that summary as review input before changing permanent instructions.

6. Sync durable artifacts across machines

When work moves between personal machines or between planning and implementation sessions, sync the durable output, not the session database itself:

Export the artifacts worth keeping:
- project instructions updates
- architecture notes
- migration checklists
- retro summaries
- reusable implementation review reports

Store them in a private repository or internal docs location after removing secrets or
machine-local details.

On the receiving machine:

1. pull the durable artifacts
2. load them as context for the new session
3. use /resume only for local session continuity, not as a cross-machine sync mechanism

session_store remains a read-only local history layer. Cross-machine continuity should come from intentionally curated docs and artifacts, not from treating session memory as a shared database.

Promotion Heuristics

Promote an insight when it is:

• repeated across multiple tasks or sessions
• specific to this repository or workflow
• likely to change future execution quality
• stable enough that you want it applied by default

Keep it as session-only context when it is still experimental, one-off, or weakly supported.

Related Skills

• cross-session-memory — search and resume prior sessions
• context-prime — load durable project guidance at session start
• ecosystem-intake — turn outside signals into adopt/adapt/reject candidates
• architecture-decisions — record major technical decisions explicitly