Skip to content

depth/ii (draft): propose intent from the data instead of asking for it - #16

Closed
sergiobuilds wants to merge 2 commits into
LilMGenius:mainfrom
sergiobuilds:pr-e-depth-ii
Closed

depth/ii (draft): propose intent from the data instead of asking for it#16
sergiobuilds wants to merge 2 commits into
LilMGenius:mainfrom
sergiobuilds:pr-e-depth-ii

Conversation

@sergiobuilds

Copy link
Copy Markdown
Contributor

Inverts the interview shape. Instead of asking the user what they want, ii reads the data they just handed over — files, a repo, a paste, a chat scrollback — and returns what the user must be trying to do with it, as a proposal the user only has to confirm, revise, or reject.

Posted as a draft — the shape is deliberately thin, so the review can settle on where this fits before it grows.

What it does

The failure mode ii targets is the well-known one: a user hands over material with a thin ask, and a follow-up question forces them to compose a sentence they did not have ready. Asking is expensive for the user; confirming is cheap. ii picks the cheap side.

Design commitments

  • Read before proposing. A proposal without reading is a guess.
  • Ground the proposal in the data. Point at what you read, not what you assumed.
  • One intent per turn. If two are plausible, propose the more likely and note the other.
  • Confirm or correct, do not compose. If the turn ends with the user needing to assemble a sentence, redo the proposal.
  • Thin-data honesty. If the data is too thin to propose from, say so and stop — do not fall back to asking.

Where it sits

Placed in depth/ because each invocation acts on one data drop, right now — the user's current handover. Contrast with the neighboring skills:

  • shower — read one artifact cold, does it stand on its own
  • hate — kill one plan with the load-bearing objection
  • ii — read one data drop, propose the intent behind it

Verification

The user's next reply is short — a yes, a correction, or a small adjustment. If they had to explain what they wanted from scratch, the proposal was wrong or was really a question.

Registered in plugin.json and README's depth/ index; passes validate-skills.sh (18 skills).

Open questions for review

  1. depth/ vs a new home. Placed in depth/ because it acts on one data drop right now. A case could be made for breadth/ (reconciling many scattered pieces into one truth — the intent).
  2. Model-invoked vs user-invoked. Draft marks it model-invoked so it fires whenever data lands with a thin ask; user-invoked would be safer but slower to help.
  3. Description length. Deliberately at the fat end of the range so the invocation trigger is unambiguous — the whole point is that it fires on ambiguous handovers.

Inverts the interview shape. Instead of asking the user what they want, ii
reads the data they just handed over — files, a repo, a paste, a chat
scrollback — and returns what the user must be trying to do with it, as a
proposal the user only has to confirm, revise, or reject.

The failure mode this skill targets is the well-known one — a user hands
over material with a thin ask, and a follow-up question forces them to
compose a sentence they did not have ready. Asking is expensive for the
user; confirming is cheap. ii picks the cheap side.

Rules keep it restrained:

- Read before proposing. A proposal without reading is a guess.
- Ground the proposal in the data — point at what you read, not what you
  assumed.
- One intent per turn. If two are plausible, propose the more likely and
  note the other.
- Confirm or correct, do not compose. If the turn ends with the user
  needing to assemble a sentence, redo the proposal.
- If the data is too thin to propose from, say so and stop — do not fall
  back to asking.

Registered in plugin.json and README's depth index; passes
validate-skills.sh (18 skills).
New skill under the depth/ perspective (minor bump per CLAUDE.md's
version-bump rule for a new skill).
@sergiobuilds
sergiobuilds marked this pull request as ready for review July 15, 2026 13:08
@sergiobuilds

Copy link
Copy Markdown
Contributor Author

Draft → Ready. Rebased onto latest `main` (past the `re0-*` rename and the new skills) and re-verified.

What the branch carries now:

  • `skills/depth/ii/SKILL.md` — unchanged from the draft. Description 324 chars; deliberately thin: four required sections (Goal / Workflow / Rules / Verification), no cross-skill name references, no `../` deep refs. Thinness is the point — the whole skill is "read the data, propose the intent, wait for confirm/correct/adjust." A longer workflow would defeat that.
  • `.claude-plugin/plugin.json` — appends `./skills/depth/ii` to the current 22-skill list; rest of the array untouched.
  • `README.md` — inserts the `ii` row into the `depth/` table (positioned right after `sip`).
  • `package.json` — bumps `0.12.0` → `0.13.0` per the version-bump rule (new skill = minor).

Verified locally: `bash scripts/validate-skills.sh` → `✓ skill catalog valid (23 skills)`, `node -e "JSON.parse(...)"` on `plugin.json` OK.

One deferred item, same as #15: `docs/readme/README.*.md` still don't list `ii` in ten languages. Held out of this PR to keep it single-purpose; happy to add a follow-up commit with the localized rows on request.

@LilMGenius

Copy link
Copy Markdown
Owner

@sergiobuilds 🏹 this filled a gap nothing else in the suite reached, thank you! Your ii skill ships in v0.15.0 as aim.

The rename: "ii" kept its secret too well, nobody half-guesses "propose the intent behind handed-over data" from two letters. "aim" says what it does, and it lands as the mirror of nba (nba proposes the next action from live state, aim proposes the intent from a data drop).

What I tuned on top of your design:

  • Rewrote it to the house SKILL.md anatomy and tightened the workflow and rules, your propose-do-not-ask core kept whole.
  • Left it model-invoked, because the reflex only earns its keep if it fires on its own the instant data lands with a thin ask.
  • Registered it for 0.15.0 across every surface and recorded the aim and nba pairing in the guide.

The idea, read the data and offer the intent to confirm instead of asking for it, is yours. Thank you 💚

@LilMGenius LilMGenius closed this Jul 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants