Use this skill after a sample or group of samples is complete, especially when
the user corrected a workflow or revealed a preference that should help future
sessions. Update documentation and memory; do not start new sample work unless
the user asks for it.
-
Locate the repository root and inspect only the relevant AGENTS.md,
docs/, wiki/, sample README.md files, notebooks, changed files, and
recent verification output. Preserve unrelated user changes.
-
Extract only durable, transferable learnings from the user's requests,
corrections, and verified results. A learning belongs in memory only when it
would help build, refactor, or troubleshoot a different sample or skill.
Separate facts from assumptions. Never record API keys, tokens, private
URLs, or other secrets.
-
Apply progressive disclosure:
- Keep
AGENTS.md short: stable project rules, conventions, and links.
- Keep
docs/ as the Pi course: concepts, learning order, and links to real
repository examples; do not duplicate complete code from samples or
notebooks.
- Keep
wiki/samples.md short: one useful paragraph per sample and a link
to its README.
- Keep detailed history in
wiki/session-memory.md.
- Keep runnable commands and implementation detail in each sample README or
notebook.
-
Update AGENTS.md with only rules that will remain useful across sessions.
Add or repair links to the course, wiki memory, and sample catalog when
needed.
-
Keep the course current. When a completed sample or notebook teaches a new
concept, update the most relevant file under docs/ or add a focused course
page. Explain the idea briefly, link to the real sample/notebook with a
repository-relative path, and mention what the learner should observe. Use
the actual artifact as the source of truth; avoid copied code that can drift.
-
Create or update wiki/session-memory.md using this structure, newest entry
first:
# Session Memory
## YYYY-MM-DD — Short topic
- Reusable principles:
- User preferences:
- Validation lessons:
- Documentation map:
Keep entries concise, factual, and linked to affected files. Record the
principle, constraint, or decision—not the story of one implementation.
Merge duplicates instead of repeating the same learning.
Do not record sample numbers, exact commands, prompt/output examples,
session IDs, timestamps beyond the entry date, temporary alternatives,
implementation filenames, or incidental model/version details. Keep those
in the sample README, code comments, test output, or course page. Link to an
artifact only in Documentation map when it helps a future maintainer find
the source of truth.
-
Update wiki/samples.md only when the sample's purpose, usage, or lesson
changed. Do not copy the full sample README into the catalog.
-
Validate with git diff --check, confirm referenced files exist, and run a
lightweight relevant check. Re-run an external model call only when it is
necessary and safe; use existing verified output otherwise.
Report the files updated, the durable learning captured, the course pages
updated, and the checks run. If the session contains no durable learning, say
so and avoid adding noise to the memory.