look-again

Look Again

Save conversation checkpoints to revisit later, and list them on demand.

Two modes. Detect from the user's phrasing:

Signal (TR / EN)	Mode
"look again", "tekrar bak" (as a command, not a question), "bunu kaydet", "buna sonra tekrar bakacağım", "save this for later"	SAVE
"tekrar bakacaklarım nelerdi", "look again list", "kayıtları göster", "what did I save", "neleri kaydetmiştim"	RECALL

If genuinely ambiguous, ask one short question. Otherwise infer.

---

Environment detection (do this first, in both modes)

• Claude Code / persistent local machine: ~/.claude/look-again/ survives between sessions. Read/write there directly.
• Claude.ai (sandboxed VM): the filesystem resets between conversations. Never pretend a write to ~/.claude/ persists.
  • SAVE → create the entry file in /mnt/user-data/outputs/ and present it to the user for download. Tell them: keep these files (e.g. in their local ~/.claude/look-again/), and upload them back to a future chat to use RECALL.
  • RECALL → look for uploaded .md entry files in /mnt/user-data/uploads/. If none are present, do NOT fabricate a list — tell the user to upload their saved look-again files (or run this in Claude Code).

Quick heuristic: if /mnt/user-data/outputs exists, you are on Claude.ai; if ~/.claude/ is a real persistent home directory (Claude Code CLI), write there.

---

SAVE mode

What gets captured

The purpose of an entry is seamless continuation: a future session reading only this file must be able to resume the work with full context, consistent terminology, and no re-asking of settled questions.

1. Context Snapshot — structured state capture (see template sections): goal, settled decisions with rationale, rejected alternatives (so they are never re-proposed), constraints & exact technical details (names, versions, paths, endpoints), current state, ordered next steps, genuinely open questions. Be specific — vague summaries break continuation. Written in the conversation's language.
2. Final Plan / Deliverable (verbatim) — if the conversation produced a final plan, design, document, schema, or code (the thing brought "to the final stage"), embed it exactly as last agreed. This — not the last message — is usually what the user wants back.
3. Last request (verbatim) — the user's most recent substantive message before the trigger. If the trigger message itself contains substantive content beyond the command (e.g. "şu hatayı çöz ... look again"), use that message minus the command phrase.
4. Last response (verbatim) — Claude's full reply to that request, exactly as written. Do not summarize, do not trim. If the response included file artifacts, list their names/paths after the verbatim text.
5. Resume Instructions — keep the template's standing block; add entry-specific notes if needed.

File format

Use the template at assets/entry-template.md. Rules:

• Filename: YYYY-MM-DD-HHmm-<topic-slug>.md (e.g. 2026-06-10-1432-clerk-jwt-validation.md). Build the slug deterministically:
  1. Take 2–5 words that name the topic.
  2. Transliterate Turkish characters: ç→c ğ→g ı→i ö→o ş→s ü→u (and uppercase equivalents).
  3. Lowercase; replace every non-alphanumeric run with a single hyphen; trim leading/trailing hyphens.
  4. Collision: if the target filename already exists (same minute, same topic), append -2, -3, … to the slug. Never overwrite an existing entry.
• Frontmatter is the index. RECALL mode reads only frontmatter, so title and summary (one line, ≤160 chars) must stand alone without opening the body.
• Context fields are mandatory. Fill project, directory, chat:
  • Claude Code: directory = pwd; project = git repo name (basename $(git rev-parse --show-toplevel) if inside a repo, else cwd basename); chat = -.
  • Claude.ai: chat = the conversation's title/topic (infer from the conversation if not explicit); project = Claude Project name if the chat is in one, else -; directory = -.
• Verbatim blocks are wrapped in five-backtick fences (`````) so nested Markdown/code fences inside the original messages don't break the file. If the verbatim content itself contains a run of five or more backticks, use a fence one backtick longer than the longest run inside the content.

Steps (Claude Code)

mkdir -p ~/.claude/look-again

Write the entry file there, then confirm to the user with the saved path and the one-line summary. Nothing else — no echo of the full content.

Steps (Claude.ai)

Write the entry file to /mnt/user-data/outputs/, present it, and remind the user to store it (ideally in their local ~/.claude/look-again/) and re-upload it when they want to recall.

---

RECALL mode

1. Collect entry files:
   • Claude Code: ls ~/.claude/look-again/*.md
   • Claude.ai: .md files in /mnt/user-data/uploads/ that contain look-again frontmatter (id: + summary: keys).
2. Read only the frontmatter of each file (e.g. awk '/^---$/{n++; if(n==2)exit} n==1' file or head -n 15). Do not load full bodies — entries can be large.
3. Present a list sorted newest-first. List language = the language of the message that triggered RECALL; if that's ambiguous, use the current conversation's language (entry language: fields describe the entries, not the list). Entries with status: blocked or status: finalized get that status appended to their summary cell (e.g. … (finalized)); they are listed, not hidden.

| # | Tarih | Başlık | Nerede | Özet |
|---|-------|--------|--------|------|
| 1 | 2026-06-10 14:32 | Clerk JWT doğrulama | cofiner-api | Clerk'in .NET SDK eksikliği; JWKS ile manuel doğrulama planı. |

("Nerede" = project if set, otherwise chat.)

4. End with: "Birini açmamı ister misin? (# numarası yeter)". When the user picks one — or asks to continue a saved topic — load that single file fully and switch into continuation mode:
   • Adopt the Context Snapshot as established working context; follow the entry's Resume Instructions.
   • Open with a 3–5 line recap (goal, where it stopped, next step) — then continue from Next Steps step 1 unless redirected.
   • Treat Decisions as settled and never re-propose Rejected Alternatives. Ask only about items listed under Open Questions, and only if blocking.
   • Keep terminology, naming, and output format consistent with the entry's Final Plan section.
5. Zero entries found → say so plainly; on Claude.ai additionally explain the upload requirement. Never invent entries.

---

Edge cases

• Trigger as the very first message (nothing to save): say there's no prior exchange yet; offer to save after the next response.
• git rev-parse fails (not a repo, git missing): set project = cwd basename; never let this abort the save.
• ~/.claude/look-again/ can't be created (permissions, read-only fs): report the error and offer to write the entry to the current directory instead — never drop the content silently.
• Reopening a finalized entry: continue normally, but note in the recap that the entry was marked finalized and ask whether to reopen it (set status: in-progress) or just consult it.
• Duplicate save of the same exchange: allowed, but note the earlier entry's filename in the confirmation.
• Very long responses: still verbatim. Disk is cheap; fidelity is the point of this skill.
• Corrupted/foreign .md in the folder: skip files lacking the id:/summary: frontmatter keys; mention skipped count only if >0.