paper-profile

name

paper-profile

description

Interactively elicits and stores the author's paper positioning so every other skill in this repo behaves context-awarely. Use when a researcher says "set up my paper profile", "ask me about my paper", "what verticals", "configure for my paper", "remember my positioning/style", or is starting a new paper and wants the toolkit tuned before drafting. Asks with sensible options about vertical/emphasis, contribution type, audience and venue tier, risk appetite, and writing preferences, then writes .paper-memory/profile.yml in the paper working directory. Explains how downstream skills consume it for benchmark-paper weighting, simulate-reviewers persona calibration, polish-prose/match-style register, and tailor-to-venue framing. Bundles a stdlib script profile_io.py that reads/writes/validates the file and emits a blank template. Local-only, never uploaded; advisory copilot, never submits.

Paper Profile

Elicits the author's paper positioning once, stores it in .paper-memory/profile.yml, and lets every other skill read it so the whole toolkit behaves context-awarely instead of asking the same questions over and over. This is the "ask me about my paper" front door: a short, optioned interview, then a small validated YAML file the rest of the repo consumes.

It is a copilot, not an oracle. The profile records your stated intent (what kind of paper this is, who it's for, how bold you want to be). It does not judge whether the science is good and it never predicts acceptance.

When to use

• "Set up my paper profile" / "ask me about my paper" / "what verticals?"
• Starting a new paper and wanting the toolkit tuned before drafting.
• "Remember my writing style / positioning across skills."
• Any other skill notices .paper-memory/profile.yml is missing and you want to create it so that skill can personalize.
• Re-run any time the positioning changes (e.g. you drop down a venue tier, or pivot from systems to empirical framing).

Inputs

• The user's paper working directory (where they want .paper-memory/ to live). This is the user's paper repo, not this skills repo.
• The user's answers to the interview (you ask; they pick). Nothing else is required — there is no network call and no file the user must pre-create.
• Optional: an existing .paper-memory/profile.yml to update instead of starting fresh.

Process

1. Locate the paper directory and the memory dir. Confirm with the user where their paper lives; the profile goes in <paper-dir>/.paper-memory/. If one already exists, load and show it (profile_io.py show) and offer to update rather than overwrite.

2. Show the option menu, then interview. Print the blank template so the user sees the choices, then ask through them. Get the exact vocabulary from the script so you never invent a value: python3 scripts/profile_io.py schema. Ask in this order, always offering the options and a one-line gloss of each (full descriptions live in references/positioning-axes.md):

   • vertical / emphasis (required): systems | theory | applied | empirical | survey | position. "Is the heart of the paper a built artifact, a proof, a domain application, a measurement study, a synthesis, or an argument?"
   • contribution_type (required): method | system | theory | dataset | empirical | application | survey | position.
   • audience: specialists | broad-field | practitioners | interdisciplinary.
   • venue_tier (required): top | specialized | regional | journal | workshop | preprint | undecided; plus any concrete target_venues (e.g. sigspatial-2026).
   • risk_appetite (required): safe | balanced | ambitious. "Defend a tight incremental delta, or stake a big claim and accept polarized reviews?"
   • writing_preferences: person (we/I/impersonal/venue-default), tone, notation (heavy/light), british_spelling; plus preferred_terms and avoid_terms.
   • context: prior_papers (paths/ids, for match-style), constraints (hard deadline, must stay anonymized, no new experiments).

   Offer a sensible default for each and let the user accept it. Do not force every field — only the four required ones must be set.

3. Write and validate. Persist the answers with one call. The script validates against the closed vocabulary and stamps the date; it refuses to write an out-of-vocabulary value (so a typo can't silently corrupt the file that every other skill trusts):

   python3 scripts/profile_io.py write <paper-dir>/.paper-memory/profile.yml \
       --field vertical=systems --field contribution_type=system \
       --field venue_tier=top --field risk_appetite=ambitious \
       --field audience=broad-field \
       --field wp.person=first-person-we --field wp.tone=assertive \
       --field target_venues="sigspatial-2026, vldb-2027" \
       --field constraints="hard deadline 2026-08-01, no new experiments"
   

   Use --from <existing> to update in place, repeated --field to set each answer, comma-separated values for list fields, and wp.<key>=... for the nested writing preferences. Re-validate any hand-edited file with python3 scripts/profile_io.py validate <path>.

4. Set up .gitignore (ask first). .paper-memory/ is local — it holds the author's private positioning and accumulated lessons. Recommend adding .paper-memory/ to the paper repo's .gitignore unless the user deliberately wants to version it (e.g. to share positioning with co-authors). State the choice; let the user decide.

5. Explain how downstream skills consume it. Tell the user concretely what changes now that the profile exists (see the table below and references/downstream-consumption.md). The point of the interview is that they won't be re-asked.

6. Mention the rest of .paper-memory/. This skill owns profile.yml. The same directory also accumulates lessons.md (deduped, dated lessons other skills append when they catch something, and read at start to avoid repeating advice) and decisions.md (venue/track/positioning decisions with rationale). See the shared .paper-memory/ convention for the file formats and memory-hygiene rules. This skill does not write those two files; it just establishes the directory and explains them.

How downstream skills consume the profile

Skill	Reads	Effect
benchmark-paper	vertical, contribution_type	Re-weights scorecard dimensions (a theory paper isn't penalized for a thin evaluation; a system paper is).
simulate-reviewers	vertical, venue_tier, risk_appetite	Calibrates reviewer personas + harshness; an ambitious claim at a top venue gets a skeptic, not a rubber stamp.
polish-prose	writing_preferences, avoid_terms	Tunes the de-AI-ify / register pass to the author's person, tone, spelling, and banned terms.
match-style	writing_preferences, prior_papers, preferred_terms	Seeds the target voice and terminology so alignment matches the author.
tailor-to-venue / select-venue	venue_tier, target_venues, contribution_type	Frames the contribution and shortlist toward the stated targets.
write-abstract	vertical, key_claim, audience	Leads with the claim the right reader cares about.

These skills should degrade gracefully: if profile.yml is absent they ask the user (or use venue defaults) as they do today. The profile removes the re-asking; it is never a hard dependency.

Output

• <paper-dir>/.paper-memory/profile.yml — a small validated YAML file (schema v1). Required: vertical, contribution_type, venue_tier, risk_appetite. Optional everything else.
• A short plain-English recap of the positioning and the concrete behavior changes in the skills the user is likely to run next.

Adapt to your discipline

The positioning axes are CS-flavored (systems/theory/empirical, conference tiers, double-blind constraints). For other fields, edit the vocabularies at the top of scripts/profile_io.py (e.g. add clinical-trial or humanities-essay verticals, swap venue tiers for journal quartiles) — the validator and emitter are data-driven, so new disciplines need new tokens, not new code. Bump SCHEMA_VERSION if you change required fields.

Guardrails

• Never invent a vocabulary value. Pull the allowed values from profile_io.py schema; the writer rejects anything else so the file every other skill trusts can't be silently corrupted.
• Never decide for the user. Offer options and a default; the author picks their own positioning and risk appetite. Don't infer "ambitious" because the topic sounds exciting.
• Local and private. .paper-memory/ is never uploaded anywhere; always offer the .gitignore line and respect the user's versioning choice.
• This skill is configuration, not authorship. It records intent. It does not write paper content, predict acceptance, or submit anything.
• Keep this file under 500 lines; references/ go one level deep only.