Enter the Conversation

The premise

Publishing is a cocktail party. A PhD student is allowed to butt into a conversation, but as a career progresses that stops working — you have to be welcomed in, and people have to be glad you showed up. The move that gets you welcomed is almost never "everything here is wrong." It is "here is the thing you were already reaching for, and I brought it."

This skill exists to find that thing. It locates the conversation, works out where its participants are currently standing, and identifies the specific opening where a contribution would be received as a gift rather than an interruption.

Two commitments shape everything below.

Be honest about the literature. The skill helps position work relative to scholars whose review you may fear. It does that by finding accurate, generous framings and by using the legitimate machinery that exists for this (field-appropriate framing, target journal choice, non-preferred-reviewer fields). It does not help omit a relevant literature to dodge its scholars. That produces a weaker paper and a worse desk-reject.

Stop at every checkpoint. Each stage produces an artifact the user reads before the next stage runs. Positioning judgments compound — a wrong read at stage 2 silently corrupts stage 7. Show the work, wait for a response, revise if asked. Do not chain stages together to save time.

Setup

Run once at the start of a session.

1. Ask for the institution. Prompt the user for their current institution and department (e.g. "University of Maryland, Robert H. Smith School of Business — Management & Organization"). Also ask whether to score the whole department or a specific group within it. Resolve to an OpenAlex institution ID with scripts/oa.py institution "<name>". Do not assume an institution from context; people move, and they also run this on target departments during a job search.

2. Set the polite-pool contact. export OPENALEX_MAILTO="you@example.edu". OpenAlex is free but throttles anonymous traffic; a contact address gets a faster, more reliable pool.

3. Create the workspace in the project directory:

conversation/
├── 00-claim.md          your claim, decomposed into concept nodes
├── 01-terms.md          term audit
├── 02-candidates.json   harvested papers with metadata
├── 03-internal.json     your internal-relevance judgments
├── 03-scored.csv        merged scores and quadrants
├── 04-home.md           home institution score and bridge map
├── 05-frontier.md       frontier read per vein
├── 06-structure.md      bifurcation, network, author and institution stats
└── REPORT.md            the assembled deliverable
papers/
├── bridge/              high internal, high external — the direct line
├── frontier/            where the conversation is live right now
├── monuments/           must-cite, low frontier signal
├── home/                the shortest bridge to each faculty member
└── papers.csv           the full scored table

papers/ holds one markdown stub per paper (metadata, abstract, why it is here, which concept node it bridges) plus PDFs if the user supplies them. Sorting a paper into a subfolder is a claim about its role, not a filing convenience — the folders are the argument.

Stage 1 — Decompose the claim

Goal: turn prose into a small set of concept nodes and the links the user is asserting between them.

Take whatever the user pasted — abstract, paragraph, half-formed pitch, chat message — and extract:

• Nodes: the 3–7 concepts the argument runs on.
• Links: the asserted relationships between nodes. These matter more than the nodes. A literature usually has good coverage of individual concepts and thin coverage of the specific joins between them; the joins are where contributions live.
• The load-bearing link: the one that, if a reviewer rejected it, would collapse the paper.
• Setting and method, if stated.

Write 00-claim.md. Checkpoint: show the user the node/link decomposition and ask whether the load-bearing link is correctly identified. People are often surprised here — the link they are defending is not always the one they think they are defending. Do not proceed until they confirm.

Stage 2 — Term audit

Goal: for each node, determine whether the literature already has a name for it.

For each concept node, search OpenAlex titles and abstracts for the user's phrasing and for plausible synonyms, then read what the top-cited hits actually mean by the term. Three outcomes, and they lead to different papers:

• Already named, same meaning. Adopt the existing term. Using a private synonym for a named construct reads as not having done the reading, and it makes the paper invisible to keyword search.
• Already named, different meaning. The most dangerous case. The term will be read against its established definition regardless of how the paper defines it. Either rename, or make the redefinition explicit and central.
• Not named. Either a genuine opening (naming it is part of the contribution) or a signal the field has been searched under vocabulary the user does not know. Check the second possibility before celebrating: search the reference lists of the closest papers for the concept under other names.

Write 01-terms.md with, per node: the user's phrasing, the established term if any, the canonical citation that fixed the definition, the outcome, and a recommendation. Checkpoint: this often changes the search vocabulary for stage 3, so confirm before harvesting.

Stage 3 — Harvest candidates

Goal: a candidate pool wide enough to contain the papers that matter and narrow enough to score.

Three complementary passes, because each one misses different things:

1. Concept search. Query title-and-abstract for each node and each established term from stage 2, and for the joins between nodes (papers containing two node terms together are the highest-value targets).
2. Citation chaining. Take 3–8 anchor papers — the ones the user already knows are central, plus the canonical citations from stage 2 — and walk both directions. Backward (referenced_works) finds the foundations. Forward (cites:) finds who is standing on them now, which is how the frontier gets found.
3. Bibliographic coupling. Papers that share many references with the anchors are working on the same problem even when they share no vocabulary. This is the pass that finds the adjacent-field work the user would otherwise miss and a reviewer would otherwise catch.

python scripts/harvest.py --config conversation/harvest.json --out conversation/02-candidates.json

Target roughly 150–400 candidates. Checkpoint: report counts by pass and show the 20 highest-cited hits. If the user does not recognize most of the top hits, the query vocabulary is wrong — fix it before scoring.

Stage 4 — Score

Goal: place every candidate on the two axes, and separate what is prestigious from what is alive.

Read references/scoring.md before this stage. The short version:

Internal relevance (0–5, judged, not computed). How much does this paper do work for the specific argument? Score it and type it:

Type	What it does
bridge	supplies a mechanism connecting two of the user's nodes
defines	fixes a term the user uses
setting	supplies the empirical context or case
method	supplies the identification strategy or analytic approach
adjacent	same neighborhood, does no work for this argument

bridge papers are the point of the exercise and the hardest to find by keyword, because they usually sit at the seam between two literatures.

External standing, split in two. Do not collapse these:

• Prestige — total citations, journal standing, author standing. Answers "will citing this be read as competent?"
• Liveness — recent citation rate and momentum, age-normalized. Answers "is anyone still standing here?"

A 1990 paper with 40,000 citations is a monument: cite it, learn nothing about the frontier from it. A 2023 paper accruing 60 citations a year is where people actually are. Collapsing them into one score makes the reading list into the syllabus the user has already read.

Write internal judgments to 03-internal.json, then merge:

python scripts/score.py --candidates conversation/02-candidates.json \
  --internal conversation/03-internal.json --out conversation/03-scored.csv

Sort papers into papers/bridge/, papers/frontier/, papers/monuments/. Checkpoint: show the 2×2 with papers placed, and specifically ask whether anything in the high-internal / low-external cell is actually good. That cell holds both the user's best unfair advantage and their worst self-indulgence, and only they can tell which.

Stage 5 — Home institution score

Goal: how far is the walk down the hall?

These are the first people the user will bring an idea to. The score answers a practical question: how much bridge-building will that conversation require?

python scripts/home.py --institution "<OpenAlex ID>" --department "<name>" \
  --candidates conversation/02-candidates.json --out conversation/04-home.md

Produce two things, and the second one matters more:

• A department aggregate — overall overlap between the department's corpus and the candidate set. Read it as difficulty, not as verdict. A low score means the idea is foreign in the building, which is a warning about the advice process, not about the idea.
• A per-faculty bridge map — for each faculty member: their closest paper to this argument, the bridge distance (0 = they are already in the candidate set; 1 = they cite or are cited by it; 2 = they share references with it; 3+ = only topical adjacency), and the one paper of theirs to cite in the meeting.

That last column is the deliverable. Walking into an advisor's office already holding the shortest bridge from their work to yours changes the conversation from "explain yourself" to "you read me."

Checkpoint: show the ranked faculty list and ask who they actually plan to talk to. Political reality is not in the citation graph.

Stage 6 — Read the frontier

Goal: where has each vein of this conversation actually gotten to?

Cluster the high-internal candidates into veins (usually 2–5). For each vein, using the liveness scores from stage 4 and the last five years of work:

• Standing question — what is currently unresolved.
• Furthest point — the most recent work that moved it, and how far.
• Stall — what has been repeated without progress. Stalls are openings; a question everyone restates and nobody answers is an invitation.
• Method ceiling — what the vein cannot currently do, and why. Often the cheapest contribution available is a method or setting that lifts the ceiling.
• Who is standing there — the 3–5 people whose next paper will be in this vein. These are the likely reviewers.

Write 05-frontier.md. Checkpoint: ask which vein the user wants to enter. Do not choose for them — the answer depends on what they want to be known for, which is not in the data.

Stage 7 — Structure of the field

Goal: the shape of the conversation, not just its contents.

python scripts/network.py --candidates conversation/02-candidates.json \
  --scored conversation/03-scored.csv --out conversation/06-structure.md \
  --graphml conversation/network.graphml

Report:

• Bifurcation. Does the citation network split into components that do not read each other? Report the split, its size, and whether the two sides use different words for the same construct. An unbridged split is the single most valuable finding this skill produces: a paper that credibly connects two camps that have stopped citing each other is welcome at the party in a way that an incremental extension never is.
• Citation network. Components, density, and the papers with highest betweenness — the existing bridges. Study how they did it; that is the genre the user is about to write in.
• Author descriptives. Most-cited, most-central, most-active-recently, rising. Flag anyone appearing in both camps of a bifurcation.
• Institutional map. Which departments own which veins, and whether the bifurcation is institutional as well as intellectual. It usually is.

Checkpoint: show the structure before writing any strategy.

Stage 8 — Entry strategy and introduction architecture

Read references/entry-strategy.md before this stage.

Produce a recommended entry: which vein, which opening (unbridged split, stall, method ceiling, or naming), and the one-sentence version of the contribution as its defenders would phrase it.

Then draft the introduction architecture — not prose, structure. The introduction decides who is in the room. Every literature named in the first three paragraphs is a literature whose scholars now expect to be cited correctly and may be assigned as reviewers; every one omitted is a potential desk-reject. Specify, paragraph by paragraph: what it does, which literature it summons, and which papers it cites.

Handle audience management explicitly and honestly. Where the user is borrowing heavily from an adjacent literature they do not want as reviewers, the answer is: cite that literature accurately and generously; frame the contribution in the target field's terms so the paper routes to the target field's editors; choose the journal accordingly; and use the submission system's non-preferred reviewer field, which exists for exactly this. Omitting the literature is not on the menu.

Assemble REPORT.md per references/report-template.md.

Bundled files

Script and reference paths are relative to the skill directory; workspace paths are relative to the user's project.

References — read the relevant one before running its stage, not after:

• references/scoring.md — full rubrics for internal relevance, prestige, liveness, quadrants, and the home institution score. Read before stage 4.
• references/data-sources.md — OpenAlex query recipes, field names, rate limits, and coverage gaps (books and working papers are the big ones, and they matter enormously in management and organization theory). Read before stage 3.
• references/entry-strategy.md — opening types, introduction architecture, reviewer and audience management. Read before stage 8.
• references/report-template.md — the exact structure of REPORT.md.

Assets — copy into the workspace and edit:

• assets/harvest.example.json — stage 3 config.
• assets/journal-tiers.example.json — optional field-specific journal weights for stage 4. Prefer these over citation-based journal proxies where the field has its own list.
• assets/roster.example.txt — faculty roster format for stage 5.

Scripts — Python 3, standard library only except network.py, which uses networkx if present and falls back to union-find components if not. Requires network access to api.openalex.org.