Claude Agents Marketplace
← Back to Skills
skillClaude Code ≥ 1.0/reflect-session

reflect-session

End-of-session capillary reflection. Extract 0-3 candidate memory entries from the session (corrections, approvals of non-obvious choices, new project facts, external references) and write them to memory/ after per-entry approval. Invoke when user signals session end (phrases like "логируем", "спасибо", "до завтра", "пока", "хватит") or when a pending_reflections/*.json marker exists. OK to output "ничего нового" and exit silently.

  • workflow

Install

~/.claude/skills/reflect-session/SKILL.md
Capillary reflection loop. Capture only what is **surprising or non-obvious** — do NOT save things already derivable from code/git/docs. If nothing warrants saving, say "ничего нового" and exit. Empty output is a valid success.

## Steps

1. **Locate source of reflection**
   - If called from pending marker: read ` (oldest first), extract `transcript_path`, read the tail of the transcript (last ~150 messages).
   - If called inline (no marker): use the current conversation context directly.

2. **Scan for signals** (classify by confidence):
   - **HIGH** → direct corrections: "не так", "no, don't", "stop", "переделай", "не мокай X". Feedback type.
   - **MEDIUM** → explicit approvals of non-obvious choices: "yes exactly", "perfect, keep doing that", "согласен" on a debatable approach. Feed
…

Paste into ~/.claude/skills/reflect-session/SKILL.md and Claude Code will pick it up on next session.

Definition

Capillary reflection loop. Capture only what is surprising or non-obvious — do NOT save things already derivable from code/git/docs. If nothing warrants saving, say "ничего нового" and exit. Empty output is a valid success.

Steps

  1. Locate source of reflection

    • If called from pending marker: read (oldest first), extracttranscript_path`, read the tail of the transcript (last ~150 messages).
    • If called inline (no marker): use the current conversation context directly.

  2. Scan for signals (classify by confidence):

    • HIGH → direct corrections: "не так", "no, don't", "stop", "переделай", "не мокай X". Feedback type.
    • MEDIUM → explicit approvals of non-obvious choices: "yes exactly", "perfect, keep doing that", "согласен" on a debatable approach. Feedback type.
    • LOW → new project facts: who/why/when, decisions, deadlines, incidents. Project type.
    • EXTERNAL → references to external systems (URLs, channels, dashboards, bot tokens). Reference type.

  3. Filter ruthlessly — skip anything that:

    • Is already in an existing memory file (Glob memory/*.md, check names/descriptions).
    • Is derivable from code, git log, or CLAUDE.md.
    • Is ephemeral (current task state, "currently debugging X").
    • Is a generic programming fact.

  4. Draft ≤3 candidates — use this exact frontmatter format: §§SHIKI0§§ Why: and How to apply: are MANDATORY for feedback and project types.

  5. Approve per entry via AskUserQuestion

    • For each candidate, one question with options: Save as-is / Edit first / Reject.
    • Show the full proposed file content in the question description or as preview.
    • If user picks Edit first, ask for the change in free text, then re-confirm.

  6. Write approved entries

    • Save to `
    • Append ONE line to MEMORY.md under the correct section: - [{name}.md](file.md) — {description}.
    • Keep MEMORY.md entries alphabetical within each section.

  7. Record session metrics — append one JSON line to ` §§SHIKI1§§

    • corrections = count of HIGH-confidence signals (user corrections/rejections in this session).
    • tasks_completed = distinct tasks worked on (estimate from conversation flow).
    • tasks_reworked = tasks where user said "не так" / rejected output and asked for redo.
    • If counts are uncertain, use best estimate — approximate data beats no data.

  8. Clean up pending markers — delete any processed pending_reflections/*.json files.

  9. Final output — 1-2 sentences: what was saved, what was skipped, plus metrics line. Example: "Сохранено 2: feedback_X, project_Y. Отклонено 1 (дубликат). Метрики: 1 коррекция, 3 задачи, 0 переделок."

Rules (hard)

  • Max 3 candidates per session. Ritual fatigue kills the loop.
  • Never auto-save. Always approve per entry (first 2 weeks of loop).
  • Never duplicate — before drafting, Glob existing memory filenames and check for overlap.
  • Never save ephemeral state. "Currently working on X" belongs in HANDOFF.md, not memory.
  • If zero signals warrant saving — output "ничего нового для памяти" and exit without asking anything.

Anti-patterns (reject these at step 4)

  • "User uses TypeScript in project X" — derivable from package.json.
  • "Fixed bug Y in commit Z" — git log owns this.
  • "User is currently debugging auth flow" — ephemeral, belongs in HANDOFF.
  • "Always use async/await" — generic programming fact.