# External Positioning > Use when a claim is about to face the outside world — classifying what is novel versus known, scoping claims to what the evidence supports, holding the figure-regeneration bar (every kept figure rebuilds from one command), and holding the reproducibility bar before anything enters a draft. Trigger phrases: "is this novel", "position against the literature", "are we over-claiming", "regenerate the figure", "can we rebuild all figures", "ready for submission", "what's our contribution". (Citation verification lives in `lit-anchor`.) - Skill: `huguryildiz/external-positioning` (Agent Skill) - Install (CLI): `npx skillmds@latest add huguryildiz/external-positioning` - Raw SKILL.md: https://api.skillmd.com/api/skills/huguryildiz/external-positioning/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Marketing & Growth - Author: huguryildiz (https://skillmd.com/u/huguryildiz) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/huguryildiz/external-positioning --- # external-positioning The project's claims will be read by skeptical strangers. This skill keeps the external story honest before it ships: novelty is bookkept rather than felt, citations are verified rather than pattern-matched, claim strength is scoped to evidence strength, and nothing enters a draft that a reviewer could not reproduce from the record. ## When to use - Writing or reviewing contribution statements, abstracts, related-work sections. - Adding any citation to the project record. - A claim is about to travel — paper draft, talk, README, funding text. - Pre-submission: the full positioning + citation + reproducibility sweep. ## When NOT to use - Internal decisions and their rationale — `decision-log`. - The statistical form of a result — `statistical-reporting`. - Verifying and formatting citations — `lit-anchor` (the sole citation authority). ## Novelty bookkeeping Keep a claims ledger (a section in the paper-planning doc or a standalone file): each paper-bound claim classified and anchored — - **novel** — no prior work found doing this; the search that failed to find it is recorded (venues/queries/date), because "we found nothing" is an auditable statement; - **known, applied** — exists elsewhere, new in this setting; cite the origin; - **confirmation / negative** — replicates or refutes something; cite what it tests. A novelty claim without a recorded search is a feeling with formatting. Re-run the search near submission — the literature moved while you worked. ## Citations Citations: see `lit-anchor` (the sole citation authority) — individual DOI resolution, distrust of bulk-generated bibliographies, the abstract-only-cannot-carry-a-load-bearing-claim rule, and tracked citation debt all live there. ## Figure regeneration **A figure that cannot be regenerated from one command does not go in the paper.** No exceptions at submission time; the time to enforce is at creation time. 1. **One command per figure.** Every kept figure regenerates via a single recorded command (a plotting script or notebook cell path + arguments) reading tracked code and evidence data — no manual pipeline of clicks. 2. **Map figure → generator → data.** Keep the mapping discoverable: a figure registry in the docs, or a naming convention (`fig3_` generated by a function/script of the same name), plus the evidence directory it reads. 3. **No hand-edited figures.** Annotations, croppings, and style fixes go into the plotting code; a figure edited in an image editor is unregenerable by construction. 4. **Regenerate on upstream change.** When the evidence data or plotting code changes, the figure is stale until re-run — treat committed stale figures as drift (`layer-sync`). 5. **Pre-submission sweep**: rebuild every paper figure from scratch and diff against the committed versions; investigate any mismatch. ## Claim scoping (honesty labels) Scope every claim to its evidence tier, in the text itself: - calibrated measurement → claim the number; - physics/theory-grounded model without calibration → claim *shape and ranking*, label the absolute values "to be calibrated"; - design operating point → say so ("chosen design point", not "optimal"); - single-setting evidence → claim the setting, not the universe ("in our configuration"). Over-claiming is the cheapest thing a reviewer can catch and the most expensive to walk back; the label costs one clause. ## The reproducibility bar (before any claim travels) A claim may enter a draft only when its chain is complete: evidence directory with meta (`evidence-convention`), environment + input + seed provenance (`run-provenance`), statistical treatment (`statistical-reporting`), regenerable figures (above), and verified citations (`lit-anchor`). The pre-submission sweep re-checks the chain end-to-end for every headline claim. ## Rules 1. **The contribution list is written against the ledger**, not from memory — every "we are the first to X" traces to a recorded search. 2. **Negative and boundary results go in the story** — a mapped dead end or a binding constraint (`negative-results-ledger`) positioned honestly reads as rigor, not weakness. 3. **Every number in the abstract exists in a table** with N and uncertainty behind it. 4. **Fix the source, then the story.** When positioning review finds an unsupported claim, the evidence gets strengthened or the claim gets weakened — the wording never bridges the gap alone. ## Configuration - `{{evidence_dir}}`, `{{decision_log}}` — where claim chains and their decisions live. ## Provenance & maintenance Authored from the scholarly-integrity canon (novelty auditing, citation verification, calibrated claim language), grounded in working practice: a one-at-a-time identifier verification protocol adopted after machine-generated references were caught fabricating entries; an explicit full-text-before-submission obligation for abstract-only sources; and shipped claims scoped as physics-grounded shape/ranking with absolute calibration labeled future work. See the pack's `examples/` directory for the worked exemplar mapping. Re-verify in your project: - Pick two random bibliography entries; resolve their identifiers and check they support their citing sentences. - `grep -rn "TODO.*cite\|abstract-only\|UNVERIFIED" docs/ | head` — citation debt is marked, and shrinking. - For one headline claim, walk the chain: draft sentence → table → evidence dir → meta → seeds. Any missing link fails the bar.