Hyperfocus Recovery
The user just surfaced from a deep work session, or got pulled away for an hour, and now has no idea where they were. Their brain went elsewhere; the context is gone. Your job is to reconstruct it from the artifacts they left behind.
When to activate
When the user says:
- "What was I doing?"
- "Where did I leave off?"
- "I lost my place"
- "Help me get back in"
- "I just came back from [X]"
Why no council
This is forensic reconstruction, not interpretation. A council call would add noise. Run inline.
Gathering pass
Run these in parallel, take what works, ignore what doesn't:
- Newest file in
.ink-and-agency/park/— a park file left by hyperfocus-exit. Check this first: it holds the head-state that no artifact can reconstruct (what they were about to try, what they already ruled out, what will silently disappear). When one exists and is recent, lead the reconstruction with it and use the rest of this pass to confirm it still matches the tree. git status— uncommitted changes (the freshest evidence of intent)git diff(working tree) — what they were actively editinggit diff --staged— what they had decided was readygit log --oneline -10— the trail of recent committed thinkinggit stash list— abandoned work- Modified files in last 4 hours —
find . -mmin -240 -type f -not -path './.git/*'(or platform equivalent) PARKED/TODO/FIXME/XXX in recently-touched files — breadcrumbs they left themselves. APARKEDcomment is deliberate and points at the exact re-entry line; treat it as the strongest signal in the file.- Open editor file — from IDE context if available
Output format
Where you were
One paragraph reconstructing the apparent goal from the evidence. Cite specific files and changes. Use phrases like "based on the uncommitted changes in [file]" — make it clear this is reconstruction, not memory.
What you were doing
A bulleted list of the active threads, ranked by recency and certainty:
- [High confidence] [Thread 1 — what file, what change, what for]
- [Medium confidence] [Thread 2]
- [Low confidence] [Thread 3 — old enough that it might not still be live]
Breadcrumbs you left yourself
Surface any PARKED / TODO / FIXME / XXX / // TEMP comments in recently-touched files, with file:line links. These are usually the most reliable signal of intent.
If a park file supplied a Do not lose list, restate it here verbatim — running processes and set env vars are the loose ends most likely to be gone by now, and the user needs to know which ones died.
Loose ends
- Uncommitted changes (count + summary)
- Stashes (if any)
- Failing tests (if known)
- Anything that looks half-done
Suggested re-entry point
Pick the smallest, most concrete next action. Same shape as task-initiation output — one literal thing, ≤30 seconds:
To get back in, do this: [Open
src/foo.tsline 47. Read the// TODO: handle empty casecomment. That's where you were.]
If the user wants to start fresh and abandon what they were doing, offer that as an alternative:
Or: if you want to drop this thread, [
git stash] and I can help you pick something new.
What NOT to do
- Do not guess at intent beyond what the artifacts support
- Do not produce a plan — just rebuild context and point at the next step
- Do not summarize for the sake of summarizing; if there's nothing to recover, say so plainly
- Do not commit anything, stash anything, or modify state — recovery is read-only
- Do not encourage or validate — the user wants their context back, not a pep talk
Host portability: tool names in this skill follow Claude Code conventions; on other hosts (Codex, opencode) map them by intent — see PORTABILITY.md.
Self-Evolve Loop
Journal ~/.ink-and-agency/learnings/hyperfocus-recovery.md (or workspace-local .ink-and-agency/ where the sandbox confines writes). Read it first; append what the run taught — SELF-EVOLVE.md.