How do I install the Session Capture skill?

Run npx skillmds@latest add adamradek-ux/session-capture in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Session Capture safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Session Capture?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Session Capture free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Session Capture?

adamradek-ux (@adamradek-ux) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Session Capture by adamradek-ux — Agent Skill

Name: Session Capture
Author: adamradek-ux

Session Capture

At the end of a work session (or on demand) it sweeps the current session and extracts four categories — decisions, corrections/lessons, new references, pending tasks — and PROPOSES writes to the right places (decisions log, lessons + memory, memory references, inbox/notes). Writes nothing without an OK (propose → OK → write). Human-in-the-loop version of the self-evolving memory pattern. Run AFTER the task is finished; the completion gate before declaring done is verify-before-done, not this skill. Triggers: "capture session", "what should I remember from this session", "log this session", and Czech aliases "zapiš session", "ulož co jsme zjistili".

adamradek-ux d5f5299 4.0 KB Updated Sep 17, 2026

File contents

Package what this session produced into durable memory — but propose, don't write. Goal: the next session should not have to rediscover what we already know. This is the human-in-the-loop version of the self-evolving memory pattern: no auto-extraction on a Stop hook, no automatic writes, no shipping the whole transcript anywhere.

When to run

At the end of a larger multi-step task, when a decision / lesson / new fact / task came out of it.
On explicit request ("capture session", "what should I remember", "zapiš session").
Not after a trivial session (one question → one answer, nothing durable) — propose nothing there, just say so.

Procedure

1. Sweep the session and sort into 4 categories

Go through what this session actually produced (not what you wish it had). For detecting corrections, use the same signals as this pack's session-evaluate hook (hooks/hook-session-evaluate.sh): "wrong / not like that / don't / redo / undo…". Sort into:

Category	What it is	Write target
Decision	A choice between options + why (what/why/status)	your decisions log (e.g. `notes/decisions.md`)
Correction / lesson	The user corrected you, or a general rule emerged	a lesson note + a `memory/feedback_*.md` entry
New reference	Durable technical fact, URL, path, mechanism	`memory/reference_*.md`
Pending task	What remains to be done (follow-up)	your inbox or today's note (e.g. `inbox/`)

Skip empty categories. Fewer and precise beats everything and vague.

2. Verify before proposing (anti-hallucination)

Every fact/URL/path in the proposal comes from this session, not from imagination. Whatever you are not sure about, mark as unverified — do not present it as fact.
Check for duplicates: does a memory/ file or lesson already cover this? Then propose an update to the existing one, not a new duplicate.

3. Present the proposal and wait for OK (human gate)

For each category list: the target path + the exact content to write. Then ask what to write. Write nothing without an explicit OK.

4. After the OK, write in exact formats

Decision → prepend a new entry at the top of your decisions log (newest first):

## YYYY-MM-DD: Short decision title
- **What:** …
- **Why:** …
- **Status:** Decided YYYY-MM-DD. …

Lesson → a short note in your notes plus a compact memory entry memory/feedback_<slug>.md (Claude Code auto-memory format):

---
name: feedback_<slug>
description: <one-line summary>
---

<the rule>

**Why:** <the reason>

**How to apply:** <when/where it kicks in>

Reference → memory/reference_<slug>.md, same frontmatter style (just the fact plus any links — no Why/How needed).

Task → append to your inbox or today's note. If none exists, offer to create one.

5. Update the index

For every new memory/ file add a one-line pointer to memory/MEMORY.md (- [Title](file.md) — one-line hook). Without it the memory will not load in the next session.

Conventions (keep)

File names kebab-case without diacritics. No {{}} placeholders — always concrete values.
Absolute dates (YYYY-MM-DD), never "today/yesterday".
Follow your project's frontmatter conventions for notes files.
After writing, you can suggest running verify-before-done as a final check.

---name: session-capturedescription: At the end of a work session (or on demand) it sweeps the current session and extracts four categories — decisions, corrections/lessons, new references, pending tasks — and PROPOSES writes to the right places (decisions log, lessons + memory, memory references, inbox/notes). Writes nothing without an OK (propose → OK → write). Human-in-the-loop version of the self-evolving memory pattern. Run AFTER the task is finished; the completion gate before declaring done is verify-before-done, not this skill. Triggers: "capture session", "what should I remember from this session", "log this session", and Czech aliases "zapiš session", "ulož co jsme zjistili".---Package what this session produced into durable memory — but **propose, don't write**.Goal: the next session should not have to rediscover what we already know. This is thehuman-in-the-loop version of the self-evolving memory pattern: no auto-extraction on aStop hook, no automatic writes, no shipping the whole transcript anywhere.## When to run- At the end of a larger multi-step task, when a decision / lesson / new fact / task came out of it.- On explicit request ("capture session", "what should I remember", "zapiš session").- **Not** after a trivial session (one question → one answer, nothing durable) — propose nothing there, just say so.## Procedure### 1. Sweep the session and sort into 4 categoriesGo through what this session actually produced (not what you wish it had). Fordetecting corrections, use the same signals as this pack's session-evaluate hook(`hooks/hook-session-evaluate.sh`): "wrong / not like that / don't / redo / undo…".Sort into:| Category | What it is | Write target ||---|---|---|| **Decision** | A choice between options + why (what/why/status) | your decisions log (e.g. `notes/decisions.md`) || **Correction / lesson** | The user corrected you, or a general rule emerged | a lesson note + a `memory/feedback_*.md` entry || **New reference** | Durable technical fact, URL, path, mechanism | `memory/reference_*.md` || **Pending task** | What remains to be done (follow-up) | your inbox or today's note (e.g. `inbox/`) |Skip empty categories. Fewer and precise beats everything and vague.### 2. Verify before proposing (anti-hallucination)- Every fact/URL/path in the proposal comes from **this session**, not from imagination. Whatever you are not sure about, mark as unverified — do not present it as fact.- Check for duplicates: does a `memory/` file or lesson already cover this? Then propose an **update** to the existing one, not a new duplicate.### 3. Present the proposal and wait for OK (human gate)For each category list: the **target path** + the **exact content** to write. Then askwhat to write. **Write nothing without an explicit OK.**### 4. After the OK, write in exact formats**Decision →** prepend a new entry at the top of your decisions log (newest first):```## YYYY-MM-DD: Short decision title- **What:** …- **Why:** …- **Status:** Decided YYYY-MM-DD. …```**Lesson →** a short note in your notes plus a compact memory entry`memory/feedback_<slug>.md` (Claude Code auto-memory format):```---name: feedback_<slug>description: <one-line summary>---<the rule>**Why:** <the reason>**How to apply:** <when/where it kicks in>```**Reference →** `memory/reference_<slug>.md`, same frontmatter style (just the factplus any links — no Why/How needed).**Task →** append to your inbox or today's note. If none exists, offer to create one.### 5. Update the indexFor every new `memory/` file add a one-line pointer to `memory/MEMORY.md`(`- [Title](file.md) — one-line hook`). Without it the memory will not load in thenext session.## Conventions (keep)- File names kebab-case without diacritics. No `{{}}` placeholders — always concrete values.- Absolute dates (YYYY-MM-DD), never "today/yesterday".- Follow your project's frontmatter conventions for notes files.- After writing, you can suggest running `verify-before-done` as a final check.

Frequently asked questions