Context Summary

context-summary

Purpose

Let Ryan clear the context window and start a fresh Claude.ai chat without burning tokens rebuilding state. Take the active tasking of the current chat — what's being worked on, what's decided, what's next, how we know we're done — and pack it into a self-contained brief that a fresh Claude instance can pick up cold.

The output is always a single fenced markdown block, ready to paste into a new Claude.ai chat.

Activation

Triggered by any of:

• "context summary" / "context-summary"
• "pack the context" / "pack this context"
• "fresh context handoff" / "context handoff"
• "summarize this session" / "summarize this session for a fresh context"
• "wrap this for a new chat" / "wrap this up for fresh context"
• "clear context — give me the handoff"

Execution

Step 1 — Identify the active tasking

Scan the chat and find the single active thread of work — the thing Ryan is currently trying to move forward. Ignore:

• Tangents that didn't pan out
• Background already captured in userMemories (unless actively relevant to the next step)
• Fully-resolved questions with no downstream impact
• Conversational filler

If there are multiple active threads, pick the most recent one and flag the others under "Other open threads" at the bottom.

Step 2 — Extract only what the next session needs

For the active tasking, pull:

• What — 1–3 sentences on what's being worked on and why
• State — what's done, what's in progress, what's blocked
• Decisions / constraints — choices already made that affect continuation (stack, approach, scope boundaries)
• Artifacts produced this session — filenames/paths/refs, not full content (unless small and critical)
• Next action — the single immediate next step
• Completion criteria — how we know the active tasking is done
• After completion — what comes next, or "session complete"

Do not invent files, paths, repos, or facts. Mark unknowns ``.

Step 3 — Strip chat artifacts

• No "as we discussed", "earlier in this chat", "I mentioned above"
• No narrative transitions
• No restating userMemories content
• Assume the receiving session has zero memory of this chat

Step 4 — Output

Above the code block, write 1–2 lines max:

• "Paste into a fresh Claude.ai chat to continue"
• Any flagged assumptions (e.g., "assumed repo: ryankolean/dtst")

Then output the full summary as a single fenced markdown block.

After the block, optionally offer: chain into chat-to-code-prompt if the next action is code execution, or workflow-lock if the session revealed a repeatable pattern worth templating.

Template

````markdown

Context Summary —

Project: Repo / location:

Active Tasking

State

• Done:
• In progress:
• Blocked on:

Key decisions / constraints

• —

Artifacts produced this session

• `` —

Next action

Completion criteria

• [ ]
• [ ]

After completion

Other open threads (if any)

• —

````

Rules

• One fenced code block. Always at the end of the response.
• Task-completion oriented. The summary exists to let the next session finish the active tasking and know it's done — not to archive the full chat.
• Token-efficient. No filler. If a field is empty, omit it entirely. Brevity over completeness.
• Self-contained. A fresh Claude instance with no memory of this chat should be able to continue from the summary alone.
• Reference artifacts, don't inline them. Filenames and paths, not full contents — unless the artifact is small (`.
• No emojis. Markdown headings, bullets, checkboxes only.
• Don't restate userMemories. Ryan's persistent context is already loaded in every session. Only include memory-adjacent facts if they are actively relevant to the next step.
• Push back on empty chats. If there's no active tasking worth packing, say so and ask what to summarize — don't generate filler.

Chaining

• Upstream: Any active working chat — planning, building, debugging, designing.
• Downstream: Pasted into a fresh Claude.ai chat.
• Adjacent: chat-to-code-prompt if the next action is to execute in Claude Code; this skill is for Claude.ai → Claude.ai continuation.
• Composes with: workflow-lock if the session revealed a reusable pattern; reminder-hub if the next action has a deadline worth a calendar entry.

Example

User: getting long in here — give me a context summary

Claude: Paste into a fresh Claude.ai chat to continue. Next action is a walkthrough prep, no repo involved.

````markdown

Context Summary — Cambourne second walkthrough prep

Project: Personal (real estate)

Active Tasking

Preparing for the second walkthrough of 366 E Cambourne St (week of May 4). Goal is to validate the as-is condition and build the case for negotiating $330K down to $315–320K.

State

• Done: Verbal agreement at $330K, as-is terms confirmed with Shelly (executor), electrical permit (2022) confirmed as panel-only
• In progress: Walkthrough checklist — HVAC, roof, foundation, plumbing, windows
• Blocked on: nothing

Key decisions / constraints

• Contingent on Waterford sale (Krysta handling agent outreach in parallel)
• Target reduction: $10–15K off $330K, justified by as-is / estate-sale terms
• Net Waterford proceeds at $365K floor only support 10% down + PMI at $330K Cambourne

Artifacts produced this session

• Cambourne Command Center React dashboard (Cormorant Garamond + warm cream palette) — local
• Devils-advocate hardening session — scored 82/100

Next action

Build the walkthrough checklist as a printable one-pager grouped by system, with "evidence for reduction" column.

Completion criteria

• [ ] Checklist printed and ready for May 4 walkthrough
• [ ] Counteroffer letter drafted, ready to send post-walkthrough

After completion

Send counteroffer → await Shelly's response → proceed to offer acceptance or walk.

````

Want me to chain into drafting the walkthrough checklist now, or save this summary as-is?

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ryankolean
• Source: ryankolean/summit-claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.