Un Claude Skill qui rédige un plan de rémunération commerciale — métrique, pay mix, quota, courbe d’accélérateurs, SPIF, déclencheur de clawback, politique de litiges — puis le chiffre en rejouant la distribution réelle d’atteinte de l’an dernier à travers la nouvelle courbe. Il indique ce que coûte le plan à trois niveaux de performance de l’entreprise, ce que gagne réellement le commercial médian face à l’OTE sur lequel il a été recruté, et quelles conditions de clawback le brouillon ne remplit pas. Il se termine par draft ou blocked. Aucun verdict ne signifie approuvé.
Le bundle est publié dans apps/web/public/artifacts/comp-plan-drafter-skill/ et contient SKILL.md ainsi que trois modèles de référence : references/1-plan-inputs-template.md (rôle, segment, OTE, quota, courbe, États où travaillent les commerciaux, budget), references/2-attainment-history-template.md (une ligne par commercial-année, départs inclus) et references/3-sample-output-format.md (le Markdown exact émis par le Skill, avec un exemple chiffré).
Quand l’utiliser
Six à dix semaines avant l’ouverture de l’année du plan, sur un plan dont le seul modèle de coût est un tableur qui suppose que tout le monde atterrit à 100 % du quota. Cette hypothèse est la raison d’être de ce Skill. L’étude AE 2026 du Bridge Group, portant sur 158 entreprises B2B, place 48 % des commerciaux au quota, les AE enterprise à 38 %, avec un ratio médian quota/OTE passé à 4,6x. Un plan chiffré sur une atteinte pleine n’est ni prudent ni agressif : il est valorisé face à une population qui n’existe pas, et l’erreur apparaît soit comme un dépassement budgétaire, soit — bien plus souvent — comme une équipe qui gagne discrètement nettement moins que l’OTE de sa lettre d’offre.
Également valable : un avenant en milieu d’année pour un segment, un rôle nouveau sans plan antérieur, et un post-mortem sur le plan actuellement en vigueur lorsque la rémunération variable a atterri loin du forecast et que personne ne sait dire si la cause était la courbe ou le quota.
Ce qui justifie le coût, c’est l’étape 3. N’importe quel outil de rémunération sait tracer une courbe ; presque aucun ne rejoue votre propre historique au niveau du commercial à travers la courbe que vous vous apprêtez à déployer. Cette rétroprojection transforme « les accélérateurs paraissent raisonnables » en « cela coûte 2,41 M à la performance de l’an dernier et 2,98 M si l’équipe gagne dix points ».
Quand NE PAS l’utiliser
Approuver ou publier un plan. Un plan de rémunération est un contrat. En Californie, le Labor Code § 2751 exige un écrit signé par l’employeur, avec un accusé de réception signé par le salarié, et il doit énoncer la méthode de calcul des commissions, y compris la politique de chargeback. Chaque draft émis par le Skill porte requires_counsel_review: true dans son en-tête, et aucun chemin ne le retire.
Calculer ou payer des commissions. Ce Skill rédige et chiffre le plan. Le calcul des versements, le traitement des litiges et la paie relèvent d’une plateforme ICM. Le Skill n’écrit nulle part.
Fixer le quota d’un commercial en particulier. Il modélise un quota pour un rôle et un segment. Attribuer un chiffre à une personne nommée est une question de couverture et de capacité : faites d’abord le découpage territorial, puis chiffrez le plan face à ce découpage.
Les fourchettes salariales de recrutement. Benchmarker le fixe et l’equity d’un poste face aux données Radford ou Pave est un autre travail, avec d’autres sources et un autre approbateur.
Les portefeuilles comptant moins d’une douzaine de commerciaux-années pleinement montés en compétence. Treize lignes forment une distribution dont on peut débattre. Six sont une anecdote à laquelle on a appliqué une fonction de percentiles, et le Skill renvoie blocked plutôt qu’un chiffre de coût qu’il ne peut pas étayer.
Mise en place
Remplissez les entrées du plan. Dans references/1-plan-inputs-template.md, définissez rôle, segment, effectif, OTE cible, pay mix, quota proposé et budget_ceiling. Laissez n’importe quel champ de courbe sur propose et le Skill rédigera cette partie ; figez ceux qui sont déjà tranchés. Renseignez market_ote_reference à partir d’une enquête que vous possédez réellement et nommez-la dans market_ote_source : l’alerte de rétention en sortie ne vaut que ce que vaut ce chiffre.
Exportez l’historique d’atteinte avec les partants dedans. Dans references/2-attainment-history-template.md, une ligne par commercial-année de l’année écoulée : quota au prorata, atteinte face à ce quota au prorata, mois de montée en compétence, date de départ. include_terminated: true est obligatoire et le Skill renvoie blocked lorsque ce champ est faux.
Décidez du plafond délibérément. Le modèle est livré sans plafond. Un plafond protège le budget contre un deal hors norme et produit de façon fiable le sandbagging qu’il était censé empêcher. Regardez le scénario haut du tableau de coûts avant de choisir, plutôt que d’hériter du réglage par défaut.
Lancez d’abord dry_run: true. Cela renvoie la distribution observée, le décompte des commerciaux pleinement montés en compétence et chaque ligne exclue avec son motif. La plupart des exports d’historique contiennent deux ou trois lignes à quota zéro ou à 400 % d’atteinte issues d’un seul deal, et vous voulez les voir avant qu’elles ne soient à l’intérieur d’un chiffre de coût.
Installez et restreignez les accès. Déposez le bundle dans ~/.claude/skills/comp-plan-drafter/ et définissez SFDC_TOKEN avec un accès en lecture sur Opportunity, User et Quota si vous tirez l’historique depuis Salesforce plutôt que d’un CSV. La lecture seule est le périmètre correct, pas une précaution.
Ce que fait réellement le skill
Deux passes, et la séparation est délibérée. La première rédige le plan : c’est le travail de jugement et il revient au modèle. La seconde rétroprojette le brouillon sur la distribution observée, et cette arithmétique tourne en code. Une fonction de versement par paliers appliquée à quarante lignes de commerciaux ne se reproduit pas d’une exécution à l’autre quand un modèle la calcule en contexte, et une discussion de rémunération s’effondre au moment où deux exécutions du même brouillon renvoient deux coûts de plan.
Le rapport de coût livre trois chiffres au lieu d’un : la distribution observée, plus et moins la bande de sensibilité. Un plan de rémunération est un instrument à effet de levier, et la grandeur utile est la pente. Un plan dont le coût bouge de 8 % sur une amplitude de vingt points d’atteinte ne dirige personne ; un plan qui bouge de 60 % est une exposition budgétaire que quelqu’un devrait accepter en connaissance de cause.
Les gains sont rapportés par décile, jamais en moyenne. Dans l’exemple chiffré de references/3-sample-output-format.md, le plan atterrit 21 % sous un budget approuvé tandis que le commercial médian gagne 154 900 face à un OTE de 200 000 — une combinaison qu’une revue purement budgétaire valide sans commentaire. Le Skill refuse aussi la solution de facilité : à 61 % d’atteinte médiane, aucun ratio quota/OTE défendable ne paie sa cible au commercial médian ; il nomme donc l’arbitrage réel (corriger la couverture, le territoire ou la montée en compétence — ou le dire franchement à l’embauche) au lieu de proposer un ajustement de taux incapable de combler l’écart.
Le contrôle de conformité produit une checklist, pas une conclusion. Trois conditions décident si un clawback survit à une contestation dans la plupart des États : le déclencheur est défini dans le document du plan avant le versement de la commission, le fait générateur est rattaché à quelque chose de véritablement réversible, et la récupération ne peut pas faire passer le commercial sous le salaire minimum applicable sur une période de paie. L’erreur de rédaction la plus fréquente est la deuxième — générer la commission au booking tout en récupérant sur un churn à douze mois — et le Skill nomme l’incohérence au lieu de rapporter une validation générique.
La réalité des coûts
Comme l’arithmétique au niveau du commercial se fait en code, le coût en tokens suit la taille du résumé et du document de plan, pas l’effectif. Un plan de 40 commerciaux revient à environ 1 à 3 USD par cycle de rédaction et de stress-test sur Claude Sonnet 5, au tarif API publié de 3 USD par million de tokens en entrée et 15 USD par million de tokens en sortie. Ce chiffre est une estimation dérivée du prix par token et de la longueur typique du document ; il varie avec la quantité de narratif demandée, pas avec la taille de l’équipe. Un cycle de conception prend six à douze exécutions à mesure que la courbe est révisée : prévoyez donc environ 20 USD pour la saison.
La comparaison qui compte n’est pas la dépense en outillage, c’est le calendrier. Un analyste RevOps qui construit les mêmes trois vues à la main — rejouer chaque commercial-année à travers une courbe candidate, refaire l’exercice à chaque révision et assembler la checklist par État — y passe deux à quatre jours par itération, et c’est pour cela que la plupart des équipes modélisent une seule courbe puis négocient à partir d’elle. Ici, chaque exécution prend quelques minutes plus une heure de lecture, ce qui fait tenir huit révisions dans la fenêtre au lieu d’une.
Face aux alternatives
QuotaPath — publie de vrais chiffres, ce qui est rare dans cette catégorie : Growth avec un forfait plateforme de 800 USD par mois incluant les cinq premiers utilisateurs, plus 50 USD par utilisateur et par mois sur le palier Premium, facturation annuelle, avec modélisation de plans, validations à plusieurs niveaux et accès API (page tarifaire de l’éditeur, vérifiée le 11/08/2026). Une organisation de 40 commerciaux tourne autour de 30 600 USD par an en Premium. Choisissez-le quand vous voulez que le plan vive dans le système qui calcule aussi les versements et route les validations. Il modélise bien les scénarios ; il ne vous dit pas que le commercial médian gagnera 77 % de l’OTE.
CaptivateIQ — rien de publié, facturation au siège sur les bénéficiaires et non sur les administrateurs, Vendr rapportant un contrat annuel médian de 36 120 USD sur 305 achats analysés. Son Compensation Builder Agent est entré en bêta limitée en mai 2026 et rédige des formules à partir de vos plans existants — c’est précisément le piège : une équipe avec quatre accélérateurs qui se chevauchent obtient de l’aide pour en construire un cinquième. Choisissez CaptivateIQ quand la gestion de la rémunération variable est le périmètre et que la structure du plan est déjà arrêtée.
Un consultant en rémunération — le titulaire honnête pour la conception du plan, et meilleur que ce Skill dans le travail politique consistant à faire accepter un plan. Il produit un bon plan par an et, en général, ne le rétroprojette pas sur votre historique au niveau du commercial, sauf si vous le lui remettez et payez l’analyse.
Le plan de l’an dernier avec les chiffres changés — la véritable référence dans la plupart des entreprises, et la raison pour laquelle les grilles d’accélérateurs dérivent pendant des années sans que personne ne chiffre cette dérive. Cela ne coûte rien, et c’est ainsi qu’une quatrième composante finit dans un plan que personne ne sait expliquer en deux phrases.
Points de vigilance
Un historique d’atteinte qui exclut les partants. L’attrition n’est pas aléatoire par rapport à l’atteinte : ceux qui atteignent peu partent, de façon disproportionnée. Un fichier limité aux survivants sous-estime le coût du plan et surestime la santé de la distribution, en même temps. Garde-fou : include_terminated est obligatoire, le Skill renvoie blocked quand ce champ est faux, et les partants entrent avec un quota au prorata et une atteinte partielle sur l’année.
Une distribution produite sous un autre quota. L’atteinte de l’an dernier reflète le quota et les territoires de l’an dernier. Garde-fou : le Skill enregistre prior_plan_quota_median et alerte lorsque le quota rédigé bouge de plus de quota_shift_tolerance_pct, en qualifiant alors le modèle de coût d’indicatif plutôt que de le présenter comme un forecast.
Un plan qui passe le contrôle budgétaire et fait partir des gens. Le rapport de coût est un instrument financier et validera volontiers un plan dont le commercial médian ne peut pas vivre. Garde-fou : le tableau des déciles est placé à côté du tableau des coûts, de sorte que le coût de rétention et le coût budgétaire figurent sur la même page et sont lus dans la même réunion.
Un SPIF permanent. Un SPIF sans date de fin n’est pas un SPIF, c’est une hausse de taux non documentée que personne ne revalide. Garde-fou : la ligne SPIF exige une date d’expiration explicite dans le fichier d’entrées, et le Skill refuse de rédiger la composante sans elle.
Prendre blocked pour un jugement sur la conception. Cela signifie que les chiffres ne sont pas fiables, pas que le plan est mauvais. Garde-fou : chaque retour blocked nomme le défaut de données précis et ce qui le lèverait, de sorte que la réponse est un export corrigé et non une refonte.
Stack
Claude — rédaction du plan, conception de la courbe, narratif des écarts de conformité ; l’arithmétique de rétroprojection tourne en code, pas en contexte
Salesforce — historique closed-won, enregistrements de quota et roster, quand le fichier d’atteinte est tiré plutôt qu’exporté à la main
Les fichiers d’entrées du plan et d’historique d’atteinte — les deux entrées qui rendent la sortie spécifique à votre organisation plutôt qu’à un modèle
Une plateforme ICM — CaptivateIQ, QuotaPath, ou ce qui calcule les versements une fois le plan rédigé approuvé et signé
---
name: comp-plan-drafter
description: Draft a sales compensation plan — metric, pay mix, quota, accelerator curve, SPIFs, clawback trigger, and dispute policy — from role, segment, and OTE inputs, then stress-test it by replaying last year's actual attainment distribution through the new curve. Reports modeled plan cost, per-decile rep earnings, and policy gaps. Emits `draft` or `blocked`, never an approved plan.
---
# Sales comp plan drafter
## When to invoke
Whenever someone is writing next year's sales compensation plan and the only cost model behind it is a spreadsheet that assumes everyone lands at 100% of quota. The canonical moment is six to ten weeks before the plan year opens, while the curve is still editable and before anything has been shown to reps. Also valid: a mid-year plan amendment for one segment, a new role whose plan has no precedent in the org, and a post-mortem run against the plan currently in market to explain why variable comp came in over or under budget.
Take a plan inputs file (`references/1-plan-inputs-template.md`), an attainment history file (`references/2-attainment-history-template.md`), and produce the Markdown plan document plus stress-test report shown in `references/3-sample-output-format.md`.
Do NOT invoke this skill for:
- **Approving or issuing a plan.** The output is a draft that goes to compensation, finance, and counsel. A sales comp plan is a contract; in California, Labor Code § 2751 requires it in writing, signed by the employer, with a signed acknowledgment of receipt from the employee. There is no verdict in this skill that means "ship it."
- **Calculating or paying commissions.** This drafts and prices the plan. Payout calculation, dispute resolution, and payroll belong to an ICM platform or the finance team, and the skill writes to nothing.
- **Setting individual quotas.** It models a quota level for a role and segment. Assigning a number to a named rep is a territory-and-capacity question — carve first, then price the plan against the carve.
- **Recruiting pay bands.** Benchmarking base and equity for a role against survey data is a different job with different sources.
- **Books with fewer than `min_history_reps` fully-ramped rep-years.** A distribution built from six survivors is not a distribution, and a cost model built on it is a guess wearing a table. The skill returns `blocked`.
## Inputs
- Required: `plan_inputs_path` — role, segment, headcount, target OTE, pay mix, proposed quota, curve shape, plan year, and the states reps work in. See `references/1-plan-inputs-template.md`.
- Required: `attainment_history_path` — one row per rep-year for the trailing plan year, including reps who left mid-year. See `references/2-attainment-history-template.md`.
- Optional: `budget_ceiling` — total variable compensation approved for the plan year, in plan currency. When set, the cost report is expressed against it rather than as a bare number.
- Optional: `sensitivity_band_pct` — default `10`. The company-wide attainment shift, in percentage points, used for the low and high cost cases.
- Optional: `dry_run` — boolean, default `false`. When `true`, the skill validates the history file and returns the observed distribution plus a data-quality report, without drafting a plan. Run this first.
## Reference files
Read both input templates before drafting anything. Without the attainment history the skill has no cost model and must not invent one.
- `references/1-plan-inputs-template.md` — the plan being designed: role, segment, OTE, pay mix, quota, curve, SPIF budget, work states, and the policy constraints that bound the draft.
- `references/2-attainment-history-template.md` — one row per rep-year: attainment percentage, prorated quota, months ramped, and termination date if any. The `include_terminated` flag is load-bearing.
- `references/3-sample-output-format.md` — the exact Markdown the skill emits, with a worked example. Downstream consumers (a finance model, a plan-document template) parse this shape.
## Method
Run in order. Steps 2 and 3 are the load-bearing split; do not merge them.
1. **Load and validate.** Parse both files. Count fully-ramped rep-years in the history. Below `min_history_reps` (default 12), stop and return `blocked` — say how many rows were found and what the floor is. If `include_terminated` is `false`, return `blocked` regardless of row count: a history containing only the reps who stayed is survivorship-biased in the one direction that matters, because low attainers leave, and it makes every plan look cheaper and every distribution look healthier than it is.
2. **Draft the plan.** This is the judgment pass and it belongs to the model. Produce: the metric paid on, the pay mix, the quota, the curve (threshold, target rate, accelerator tiers and where they kick in, decelerator if any), a SPIF budget line with a named expiry date, the clawback trigger, and the dispute-resolution policy with a named response window. Keep the component count at or below `max_components` (default 3) — every component past the third divides rep attention without adding steering, and the modeled dollars usually show one component carrying almost nothing.
3. **Back-cast the draft against the observed distribution — in code.** Replay each historical rep's actual attainment percentage through the *new* curve and sum the modeled payouts. Do this arithmetic in code, not by reading the roster into the reasoning context. A piecewise payout function applied to forty rep-rows will not reproduce run to run when a model does it in context, and a compensation conversation collapses the moment two runs of the same draft produce two plan costs. The model's job is the curve, the ranking, and the narrative.
Report three numbers, not one: modeled cost at the observed distribution, at the distribution shifted up by `sensitivity_band_pct`, and shifted down by the same. A comp plan is a leveraged instrument and the useful figure is the slope. A plan whose cost moves 8% across a 20-point attainment swing is under-leveraged and will not change behavior; one that moves 60% is a budget risk somebody should agree to on purpose.
4. **Report earnings by decile, not by average.** Average earnings hide the plan's actual behavior. Emit modeled total earnings at the 10th, 50th, and 90th percentile of the observed distribution, alongside target OTE. The number that predicts attrition is what the median rep actually earns against the OTE they were recruited on — if that lands well below target, the plan is priced for the budget rather than for the market and the cost report will still pass.
5. **Run the policy check.** For each state in `rep_work_states`, note the local constraint and emit the checklist rather than a conclusion. Three conditions decide whether a clawback survives challenge in most states: the trigger is defined in the plan document before the commission is paid, the earning event is tied to something genuinely reversible, and the recovery mechanism cannot push the rep below the applicable minimum wage in any pay period. Check the draft against all three and name which one fails. Where California is in the list, add the § 2751 items: written agreement, signed by the employer, signed acknowledgment of receipt from the employee, and a stated method for computing commissions including the chargeback policy.
6. **Emit `draft` or `blocked`.** `draft` means the plan is modeled, priced, and ready for human review; every `draft` document carries `requires_counsel_review: true` in its header. `blocked` means a data problem makes the cost numbers untrustworthy — too few rep-years, terminated reps excluded, or a quota shift that invalidates the distribution. There is deliberately no third verdict. The skill makes the cost and the policy gaps visible before people decide; it does not decide.
## Output format
The skill emits Markdown in exactly this shape. Full worked example with populated rows in `references/3-sample-output-format.md`.
```markdown
# Comp plan draft — FY27 Mid-Market AE
plan_year: 2027 | headcount: 34 | history rows: 41 (incl. 9 terminated)
requires_counsel_review: true
## Verdict: draft
Curve lands 21% under budget at the observed distribution and the median rep
earns 77% of OTE — priced for the budget, not the market. Clawback trigger
fails the reversibility test.
## 1. Plan structure
| component | metric | weight | notes |
|---|---|---|---|
| base | — | 50% of OTE | 100,000 |
| commission | closed-won ARR | 45% of OTE | 9.4% of ARR at target |
| SPIF | new-logo multi-year | 5% of OTE | expires 2027-06-30 |
## 2. Curve
| band | attainment | rate | cumulative payout |
|---|---|---|---|
| threshold | 0-50% | 0% | 0 |
| target | 50-100% | 9.375% of ARR | 90,000 at 100% |
| accelerator 1 | 100-130% | 1.5x base rate | 130,500 at 130% |
| accelerator 2 | 130%+ | 2.0x base rate | uncapped |
## 3. Modeled cost (back-cast on 41 rep-years, 34 plan heads)
| case | company attainment | variable cost | vs budget (3,060,000) |
|---|---|---|---|
| low | observed -10 pts | 1,940,000 | -37% |
| observed | 61% median | 2,410,000 | -21% |
| high | observed +10 pts | 2,980,000 | -3% |
## 4. Rep earnings by decile
| percentile | attainment | modeled earnings | vs OTE (200,000) |
|---|---|---|---|
| p10 | 34% | 100,000 | 50% |
| p50 | 61% | 154,900 | 77% |
| p90 | 141% | 250,300 | 125% |
## 5. Policy check
| item | state | status |
|---|---|---|
| written + signed + acknowledged | CA | present in draft |
| computation method stated | CA | present in draft |
| clawback trigger pre-defined | all | present in draft |
| clawback tied to reversible event | all | FAIL — 12-month churn is not the earning event |
| minimum-wage floor per pay period | CA, NY, WA | not modeled — needs draw schedule |
## 6. What to change
- The clawback recovers on churn inside 12 months, but the plan earns
commission on booking. Tie recovery to non-payment or contract
cancellation, or move the earning event to cash collected.
- Median rep earns 154,900 against a 200,000 OTE, and the plan spends
650,000 under budget. The curve is not the cause: at 61% median
attainment, no defensible quota-to-OTE ratio pays the median rep target.
Quota would have to fall to roughly 590,000 — a 3.0x ratio — to put the
median at 100%. Either fix the input (coverage, territory, ramp) or
accept that this plan pays half the team 77% of OTE and say so at hire.
- Cost moves 43% across a 20-point attainment swing. That is real leverage
and it is worth confirming on purpose rather than discovering in Q3.
```
## Watch-outs
- **An attainment history that excludes reps who left.** Low attainers leave, so a survivors-only file understates plan cost and overstates the health of the distribution — in the same direction, at the same time. Guard: `include_terminated` is a required field and the skill returns `blocked` when it is `false`; terminated reps enter with prorated quota and partial-year attainment.
- **A distribution produced under a different quota.** Last year's attainment reflects last year's quota and last year's territories. If the new median quota moves materially, the old distribution stops predicting anything. Guard: the skill records `prior_plan_quota_median` and warns when the drafted quota moves more than `quota_shift_tolerance_pct` (default 15), stating in the report that the cost model is directional only.
- **A plan that passes the budget check and loses people.** The cost report is a finance instrument and it will happily approve a plan the median rep cannot live on. Guard: the decile table sits next to the cost table and reports median modeled earnings as a percentage of target OTE, so the retention cost is on the same page as the budget cost.
- **A clawback that recovers against something the earning event does not cover.** Recovering commission on 12-month churn when the plan earns on booking is the most common drafting error, and it is the condition that fails in dispute. Guard: step 5 tests the trigger against the earning event explicitly and names the mismatch rather than reporting a generic pass.
- **A permanent SPIF.** A SPIF that never expires is not a SPIF, it is an undocumented rate increase that nobody re-approves. Guard: the SPIF line requires an explicit expiry date in the inputs file, and the skill refuses to draft a SPIF component without one.
- **Treating `blocked` as a verdict on the plan.** `blocked` says the numbers cannot be trusted, not that the design is wrong. Guard: every `blocked` return names the specific data defect and what would clear it, so the response is a data fix rather than a redesign.
# Plan inputs
Replace the example values with yours. Everything here describes the plan you want drafted and the constraints it has to live inside. The skill drafts the curve; it does not invent the OTE, the headcount, or the states your reps work in.
## Role and scope
```yaml
plan_year: 2027
plan_start: 2027-01-01
role: "Mid-Market Account Executive"
segment: "mid-market" # smb | mid-market | enterprise | strategic
headcount: 34 # quota-carrying heads in this plan at plan start
currency: USD
plan_status: draft # draft | socialized | in_market
```
`plan_status` changes the tone of the report, not the arithmetic. Once a plan is `socialized`, a recommendation to change the curve carries a communication cost as well as an effort cost, and the report says so in the header rather than leaving the leader to discover it in the meeting.
## Target compensation
```yaml
target_ote: 200000
pay_mix: # must sum to 100
base_pct: 50
variable_pct: 50
market_ote_reference: 200000 # what you believe the market pays this role
market_ote_source: "Bridge Group 2026 AE Metrics, median OTE 200,000"
```
`market_ote_reference` is what the decile table compares against when it flags a plan priced below market. Set it from a survey you actually hold, and name the source — if it is a guess, say so in `market_ote_source`, because the retention flag is only as good as this number.
## Quota
```yaml
proposed_quota: 960000 # annual, per fully-ramped rep
prior_plan_quota_median: 850000 # last year's median assigned quota
quota_shift_tolerance_pct: 15 # warn above this much movement
```
The tolerance exists because the attainment history you supply was produced under `prior_plan_quota_median`. Move the quota far enough and the distribution stops predicting the new plan's cost. The skill will still model it; it will label the output directional.
## Curve
Give the shape you want drafted. Leave any field as `propose` and the skill drafts that piece; pin the ones that are already decided.
```yaml
curve:
threshold_pct: 50 # no commission below this attainment
target_rate: 1.0 # multiplier between threshold and 100%
accelerators:
- from_pct: 100
to_pct: 130
rate: propose
- from_pct: 130
to_pct: null # null = uncapped
rate: propose
decelerator: none # none | {below_pct, rate}
cap: none # none | a dollar figure
```
Uncapped is a real decision, not a default. A cap protects the budget against a single outsized deal and reliably produces the sandbagging it was written to prevent; the cost table's high case is where you should look before choosing.
## Components
Keep this at three or fewer. A fourth component almost always shows up in the modeled-dollars column carrying a rounding error's worth of pay and a meaningful share of rep attention.
```yaml
max_components: 3
components:
- name: commission
metric: closed_won_arr # the thing you want more of
weight_pct: 90 # share of the variable half
- name: new_logo_spif
metric: new_logo_multiyear
weight_pct: 10
expiry: 2027-06-30 # required — a SPIF with no expiry is a rate increase
```
## Earning event and clawback
```yaml
earning_event: booking # booking | invoiced | cash_collected
clawback:
trigger: "customer non-payment within 90 days of invoice"
recovery_method: "offset against future commission, max 25% per pay period"
post_termination: false
```
`earning_event` and `clawback.trigger` have to describe the same thing. Earning on `booking` while recovering on churn is the mismatch the policy check is looking for: the plan is trying to reverse something the earning event never depended on.
## Policy constraints
```yaml
rep_work_states: [CA, NY, TX, WA, IL]
draw:
type: none # none | recoverable | non_recoverable
amount_monthly: 0
dispute_policy:
response_window_days: 15
escalation: "RevOps → VP Sales → CFO"
budget_ceiling: 3060000 # total approved variable comp for the plan year
min_history_reps: 12
sensitivity_band_pct: 10
```
The state list drives the policy checklist. It is not legal advice and the skill does not pretend otherwise — it produces the items counsel needs to see, marked against the draft, so the review is a review rather than a discovery exercise.
# Attainment history
One row per rep-year for the trailing plan year. This file is the cost model. Everything the stress test reports comes from replaying these rows through the drafted curve, so the quality of this file sets the quality of the output.
## The one rule that matters
```yaml
include_terminated: true
```
Include the reps who left. All of them, with their partial-year attainment and their prorated quota.
This is not a completeness preference. Attrition in a sales org is not random with respect to attainment — low attainers leave, and they leave disproportionately. A history containing only the people still on the roster understates what the plan will cost and overstates how healthy the distribution is, in the same direction, at the same time. The skill returns `blocked` when this flag is `false`, because a plan priced on survivors is priced on the wrong population.
## Rows
```csv
rep_id,segment,months_ramped,prorated_quota,attainment_pct,terminated_on,notes
r-001,mid-market,12,850000,141,,
r-002,mid-market,12,850000,118,,
r-003,mid-market,12,850000,104,,
r-004,mid-market,12,850000,97,,
r-005,mid-market,12,850000,88,,
r-006,mid-market,12,850000,74,,
r-007,mid-market,12,850000,61,,
r-008,mid-market,12,850000,58,,
r-009,mid-market,12,850000,44,,
r-010,mid-market,12,850000,31,,
r-011,mid-market,7,495833,38,2026-07-31,involuntary
r-012,mid-market,5,354167,22,2026-05-29,involuntary
r-013,mid-market,9,637500,96,2026-09-30,voluntary — competitor offer
```
| Column | What goes in it |
|---|---|
| `rep_id` | Any stable identifier. Do not use names; the output is circulated. |
| `segment` | Must match a segment in the plan inputs, or the row is excluded and counted in the data-quality report. |
| `months_ramped` | Months at full productivity during the year. A rep in month three of a six-month ramp contributes a partial rep-year and is excluded from the fully-ramped count. |
| `prorated_quota` | The quota actually carried, prorated for partial years. Not the annual number. |
| `attainment_pct` | Attainment against `prorated_quota`, as a whole number. `141` means 141%. |
| `terminated_on` | ISO date, blank if still employed. |
| `notes` | Free text. `voluntary` / `involuntary` is worth recording — it is the only signal in this file about whether the plan drove the exit. |
## What counts toward `min_history_reps`
Only fully-ramped rep-years — rows where `months_ramped` equals 12, or where a terminated rep was fully ramped for the months they worked. Ramping reps are still worth including for the cost model, because they cost money, but they do not make the distribution more trustworthy and the skill does not count them toward the floor.
Below the floor, the correct output is `blocked`. Thirteen rows is a distribution you can argue about; six is an anecdote with a percentile function applied to it.
## Optional: prior-plan curve
Supply the curve these attainments were paid under and the report adds a year-over-year comparison — what the same rep-years would have cost under the old plan against the new one. This is the single most persuasive number in a comp review, and it is unavailable without this block.
```yaml
prior_curve:
threshold_pct: 60
target_rate: 1.0
accelerators:
- from_pct: 100
to_pct: null
rate: 1.4
cap: none
```
## Data-quality report
Run the skill with `dry_run: true` before drafting anything. It returns the observed distribution, the fully-ramped count, and a list of rows it had to exclude with the reason for each. A history file usually has two or three rows with a quota of zero or an attainment above 400% from a single outsized deal, and you want to see those before they are inside a cost number rather than after.
# Sample output
The exact Markdown the skill emits, populated from the example inputs in `references/1-plan-inputs-template.md` and the example history in `references/2-attainment-history-template.md`. Section order and heading text are stable — a finance model or a plan-document template can parse against them.
---
# Comp plan draft — FY27 Mid-Market AE
plan_year: 2027 | headcount: 34 | history rows: 41 (incl. 9 terminated)
quota shift vs prior plan: +12.9% (within 15% tolerance)
requires_counsel_review: true
## Verdict: draft
Curve lands 21% under budget at the observed distribution and the median rep earns 77% of OTE. The plan is priced for the budget, not the market. The clawback trigger fails the reversibility test and must be redrafted before counsel review.
## 1. Plan structure
| component | metric | weight | notes |
|---|---|---|---|
| base | — | 50% of OTE | 100,000 |
| commission | closed-won ARR | 45% of OTE | 9.375% of ARR, 90,000 at 100% |
| SPIF | new-logo multi-year | 5% of OTE | 10,000 pool, expires 2027-06-30 |
Earning event: `booking`. Draw: none. Dispute window: 15 days, escalating RevOps → VP Sales → CFO.
## 2. Curve
| band | attainment | rate | payout at band ceiling |
|---|---|---|---|
| threshold | 0-50% | 0% | 0 |
| target | 50-100% | 9.375% of ARR | 90,000 |
| accelerator 1 | 100-130% | 1.5x base rate | 130,500 |
| accelerator 2 | 130%+ | 2.0x base rate | uncapped |
The threshold is a gate, not a ramp: a rep who clears 50% is paid on every dollar from the first, and a rep at 49% is paid nothing variable. That is a deliberate cliff and it is the single most disputed line in any plan that has one — the 49% rep and the 51% rep are 20,000 apart on a difference of one deal.
## 3. Modeled cost — back-cast on 41 rep-years, 34 plan heads
| case | company attainment | variable cost | vs budget (3,060,000) |
|---|---|---|---|
| low | observed -10 pts | 1,940,000 | -37% |
| observed | 61% median | 2,410,000 | -21% |
| high | observed +10 pts | 2,980,000 | -3% |
Cost moves 43% across the 20-point band. Under the prior curve (60% threshold, 1.4x single accelerator) the same 41 rep-years would have cost 2,265,000 — the new plan is 6.4% more expensive at identical performance, almost entirely from the second accelerator tier.
## 4. Rep earnings by decile
| percentile | attainment | modeled earnings | vs OTE (200,000) |
|---|---|---|---|
| p10 | 34% | 100,000 | 50% |
| p50 | 61% | 154,900 | 77% |
| p90 | 141% | 250,300 | 125% |
Nine of the 41 rep-years land below the 50% threshold and earn base only. Three of those nine are terminated rows, which is the expected pattern and the reason the terminated set has to be in the file.
## 5. Policy check
| item | state | status |
|---|---|---|
| written agreement, signed by employer | CA | present in draft |
| signed acknowledgment of receipt | CA | present in draft |
| method of computing commissions stated | CA | present in draft |
| chargeback policy stated in agreement | CA | present in draft |
| clawback trigger defined before payment | all | present in draft |
| clawback tied to a reversible earning event | all | **FAIL** — see below |
| recovery cannot breach minimum wage in a pay period | CA, NY, WA | not modeled — no draw schedule supplied |
| post-termination recovery | all | disabled in draft (`post_termination: false`) |
This is a checklist for counsel, not a legal opinion. It reports which items the draft addresses and which it does not.
## 6. What to change
- **Clawback reversibility.** The inputs file sets `earning_event: booking` and a clawback trigger of customer non-payment within 90 days of invoice. Those describe different events: the commission is earned when the deal is booked, and non-payment happens downstream of an earning event that never depended on collection. Either move `earning_event` to `cash_collected`, or narrow the trigger to contract cancellation before invoice. As drafted this is the condition most likely to fail if a recovery is challenged.
- **The median rep earns 154,900 against a 200,000 OTE**, while the plan spends 650,000 less than the approved budget. The curve is not the cause. At a 61% median attainment, no defensible quota-to-OTE ratio pays the median rep target — quota would have to drop to roughly 590,000, a 3.0x ratio, to put the median at 100%. The proposed 960,000 quota against a 200,000 OTE is 4.8x. Fix the input (coverage, territory, ramp) or accept that this plan pays half the team 77% of OTE and say so at hire rather than in month nine.
- **Minimum-wage floor is unmodeled** because no draw was supplied. With `draw.type: none` the base alone clears the floor in every listed state, so this is informational — but if a recoverable draw is added later, the 25%-per-pay-period recovery cap has to be re-checked against it.
- **Cost leverage of 43% across 20 points** is a real exposure in a good year. It is defensible; it should be a decision somebody makes in December rather than a surprise in Q3.
