Recap State

Recap

What this skill does

When you are mid-conversation and lose the thread — because of a break, a long tool output, or an agent question you can no longer place — this skill stops, restructures the last N minutes into a fixed 7-block schema, and ends with a Synthesis-check question. You read yourself back into state in under 60 seconds, confirm the next step, and then the work continues.

When to fire

Fire when you recognise any of the trigger phrases listed in the frontmatter description, or when the user explicitly says something like "where are we" / "我不知道你在說什麼" / "今どこまで来た".

This skill is in-session re-orientation. The built-in /recap (away-summary) handles the cross-context return case — user was away, comes back, gets a summary. That is different. This skill fires while the user is already in conversation and has lost the thread. The two coexist; do not replace the built-in away-summary.

What to do

1. Read references/seven-block-schema.md for the full V1 template and the

5 共通核心原則 definitions.

1. Structure your in-chat response with two-mode XML tag separation.

Output exactly two sibling top-level tags — one for private planning, one for user-facing explanation:

```

[Internal planning. Telegraphic / structured / agent-to-self mode is fine here. The user sees this block but treats it as your scratch pad, not as the recap content. Plan:

• Which user phrases from this session are spec-critical (file paths,

error messages, named constraints, exact tool/command names) — these MUST appear verbatim inside Block 2 Background.

• Which blocks would compress better as a table (per seven-block-schema.md

§1.5 — Block 3 Assessment defaults to 2-col key:value; Block 5 if 2+ options compared; Block 6 if items have metadata).

• The agent's most recent question (drives Block 5 Why-this-question).
• The expected next-step direction (drives Block 7 Synthesis-check).]

[User-facing explanation. SWITCH to conversational, second-person mode. Write as if speaking TO the user. Apply the plain-language principle in full — explanation-style not report-style. Produce 6 blocks in order below.]

### Block 1 — Situation ### Block 2 — Background ### Block 3 — Assessment ### Block 5 — Why-this-question ### Block 6 — Pending ### Block 7 — Synthesis-check

```

(Block 4 is L2-only and skipped at L3 — see references/seven-block-schema.md §"Block ↔ Audience map".)

Why two tags, not one stream: the ` block holds planning (telegraphic / structured / agent-to-self — fine to be report-style). The block holds the user-facing explanation (must be plain-language, conversational, explain-TO-user). Mixing them in one stream causes the user-facing text to inherit thinking-mode terseness and status-report tone — exactly the failure that plain-language (v0.1.1) was reframed to prevent. Anthropic's official prompting docs recommend XML tag separation ( + pattern) for precisely this kind of internal-vs-external mode boundary; this skill adopts that pattern as the structural enforcer for plain-language`.

1. Block details (each rendered inside ``):

• Block 1 — Situation: one sentence, present moment, where we are stuck.
• Block 2 — Background: 3–5 bullets; key decisions + rejected options +

critical file names / error strings. Quote directly — do not rewrite. This is where spec-critical user phrases land at L3 (since Block 4 does not render).

• Block 3 — Assessment: current assumption + confidence level + what is

unknown or blocking.

• *(Block 4 — User messages — skipped at L3. L2-only; the future

HANDOFF sister skill renders it verbatim for a cold AI reader. At L3 the human reader already remembers their own turns; dumping them adds no signal.)*

• Block 5 — Why-this-question: explain the agent's most recent question:

what is being asked, why it matters, what options the user has.

• Block 6 — Pending: checklist of tasks not yet done.
• Block 7 — Synthesis-check: one sentence stating the expected next step,

then ask the user to confirm or redirect.

1. Apply all 5 共通核心原則 throughout (4 of the 5 fire at L3; the 5th

all-user-messages is L2-only and dormant here — kept named in the principle list as the canonical SSOT for the future HANDOFF sister skill):

• structured-schema: use the fixed block structure every time. No

free-form paragraphs substituted for blocks. At L3 = 6 blocks (skip Block 4).

• quote-not-paraphrase: block 2 (and block 4 at L2) reproduces original

strings exactly — file paths, error messages, command names, user wording. At L3 this principle does the work Block 4 used to do — spec-critical user phrases land in Block 2's quote-not-paraphrase enforcement.

• all-user-messages (L2-only — dormant at L3): at L2 HANDOFF block 4

lists every user turn verbatim. At L3 (this skill) the principle does not fire; Block 4 is skipped entirely. Anchor preserved here as SSOT for HANDOFF.

• synthesis-check: block 7 always ends with a directed question; agent

does not continue until user responds.

• plain-language: each block explains TO the user in plain everyday

language — write as if speaking to them ("you were debugging X" / "we picked Y because Z"), NOT reporting at them ("Current state: X. Decisions: Y, rationale: Z."). Use second-person framing where natural. Only use a technical term if the user introduced it in this session. Expand acronyms on first mention. Short sentences. One fact per bullet. See references/seven-block-schema.md §plain-language for the report-style vs explanation-style failure-mode contrast.

1. Visual aids (tables / ASCII diagrams): permitted **only when they

compress information**, not as decoration. See references/seven-block-schema.md §1.5 for per-block guidance. Block 3 (Assessment) defaults to a 2-col key:value table. Block 5 (Why-this-question) and Block 6 (Pending) use tables when there are 2+ options or per-item metadata, prose otherwise. Skip tables for blocks 1 + 7 (single sentence). Reach for ASCII only when real topology exists (pipeline / dependency / state machine).

1. After producing the 6 L3 blocks (Block 4 skipped), wait. Do not continue

until the user responds to the Synthesis-check question in block 7.

Soft-gate Synthesis-check

Block 7 is a soft gate: any "yes / 對 / go / continue / 繼続 / はい / proceed" counts as confirmation and the agent continues. The agent does not hard-block on a specific keyword — any clear confirmation from the user is enough. If the user redirects instead of confirming, follow the redirect.

The Synthesis-check is not a formality. Ending the recap and immediately continuing without waiting is the failure mode this block exists to prevent — the agent would resume from its own assumption of alignment, and the user's correction would only surface 3 turns later.

What NOT to do

• No file output: the 7-block recap goes in chat only. Do not write a file.
• No cross-session work: this skill covers the current session only.

Cross-session continuation is HANDOFF (a future sister skill).

• Do not replace the built-in /recap: the away-summary is a different

tool for a different moment. This skill fills the in-session gap.

• Do not produce a free-form summary instead of the 6 L3 blocks: the schema is

the user-visible value. A paragraph that "covers the same ground" is not equivalent — it breaks the structured-schema principle.

• Do not render Block 4 at L3: that block is L2-only (for the future

HANDOFF sister skill's cold AI reader). At L3 the human reader already remembers their own turns. If you find yourself listing user messages, stop and ask whether the content actually belongs in Block 2 Background.

• Do not paraphrase spec-critical phrases in Block 2: file paths, error

messages, named constraints, exact tool / command names stay verbatim. At L3 Block 2 is the only place spec-critical user phrases land (since Block 4 is skipped) — losing those quotes here is invisible intent loss.

• If the user actually asks for verbatim user-message listing at L3:

produce it as a separate response outside the recap schema, not as a ressurected Block 4. The 6-block recap stays focused.

• Do not add decorative tables / ASCII: a 1-row table, a 2-row table whose

rows are unrelated, or an ASCII chart for a 3-item list adds framing without compressing. Per references/seven-block-schema.md §1.5, tables fire only when they flatten ≥3 sub-items or compare ≥2 options.

See also

• references/seven-block-schema.md — SSOT for the V1 7-block template,

5 共通核心原則 definitions, good-example / bad-example pair.

Source & license

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

• Author: kouko
• Source: kouko/monkey-skills
• License: MIT

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