ground-and-draft
Pull real customer evidence from the Amdahl corpus, freeze it, and draft GTM copy grounded in it. Use at the start of any GTM writing task — outbound email, sequence step, one-pager, call prep, websit
SKILL.mdskills/ground-and-draft/SKILL.md
GitHubground-and-draft
The first half of the loop: Write → Ground → Draft → (grade-and-report).
This skill stops at a readable draft. Do not chain straight into grading. The gap between the two skills is where a human looks at the draft and decides whether they agree with it, and that is where the judgment in this loop lives. Automate the gap away and the exercise stops producing anything you could not get from a load test.
When it applies
Any time you are about to write GTM copy and want it grounded in what buyers actually said rather than in what the model assumes they would say.
Steps
Write the intent — one line: what you are drafting and for whom (persona, segment, and a real trigger or signal if this is outbound). Vague intent is what produces a draft that grades as "not commercial outreach".
Ground — scoped the way the grade will be scoped.
export AMDAHL_KEY='...' # Settings, then Developer, Create key bash scripts/ground.sh semantic "their pain, in plain language" \ evidence/$(date +%F)-launch.json 25 \ '[{"field":"speaker_type","op":"eq","value":"external"}]'speaker_type: externalbelongs on nearly every pull. It is step 3 below, enforced at retrieval instead of eyeballed after.- Account scoping is
company_id(eq/in), and only when the account is already in your corpus. - On a cold account there is correctly nothing narrower than the cohort.
A company that has never spoken to you has no words to retrieve; a
company_idfilter would return an empty pull, and an empty pull is not grounding. Pull the cohort instead and attribute at that scope when you write — "the eng leaders we talk to say" is supported, "your engineers said" is not. That is the normal case for first-touch outbound, not a degraded one. - Audience predicates need
mode filter. The semantic lane admits onlycompany_id,occurred_atandspeaker_type;role_level,speaker_title,inferred_personaandindustryare refused there withinvalid_argument. Run them as a secondfilter-mode pull into its own dated file — rows by predicate, complementing the similarity pull rather than replacing it.GET /search/fieldslists the vocabulary per surface. - Prefer
semantic(vector similarity, fast). Thefuzzy/NL lane has a ~15s ceiling and returnsstatus=failedon broad asks. - A new pull is a new dated file. The script refuses to overwrite one.
- Scraped or competitor material is not customer evidence. Keep it in a separate directory and never feed it in as though a customer said it.
Check who is talking. The script prints a
speakersbreakdown. Unfiltered corpus retrieval does not filter on speaker side, and a meaningful share of a typical pull is your own team, not a buyer. A claim about what customers want that is grounded in your own rep's words is the quiet failure here — and it will still score, because the words are real and on-topic. Passingspeaker_type: externalin step 2 is the structural version of this check; read the breakdown anyway, because a pull that comes back thin after filtering is telling you the theme is yours, not theirs.Draft. For an outbound email or LinkedIn message, ground, then Amdahl writes. Do not draft it yourself: the Optimizer writes from context and does not read a draft. Call
POST /messages/optimize(MCP:optimize_message) with:context.facts: the recipient, the trigger or signal, what you offer, and the proof point, one fact per entry, each stated exactly as its source does. Never put a customer quote in a fact: a fact is graded as yours, not a customer's.context.rules: the artifact's length, voice and ask."evidence": "workspace", so the Optimizer reads your customers' words itself and checks the message's claims against them.
The returned
messageis the draft. Use it as returned. See https://docs.amdahl.ai/concepts/context for what counts as enough context.For any other copy (a one-pager, call prep, website copy), draft it yourself. Their problem first, in their words. One proof point. No invented numbers, no invented peer names, no invented case studies — the eval checks specifics against the corpus and a fabricated one is worse than a vague one. Obey the artifact's own length and voice rules.
Create the run folder.
runs/$(date +%F-%H%M)-<slug>/outputs/draft.mdRecord the evidence file and the prompt version in a
manifest.jsonbeside it. You will want to know which evidence produced which draft when the scorecards start piling up.
Hand the draft to a human. Then run grade-and-report.
Freeze the evidence when you are comparing
Live search is right when you are producing work. Frozen evidence is required when you are comparing two versions, because a second search returns different quotes and any difference you measure could be retrieval rather than your edit.
Two ways to freeze:
- Across drafting attempts — the dated file this skill writes.
- Across graded runs — pass
evidence_from_run: "<run id>"toevals.run, and the new run reuses the first run's quotes with no file to manage.grade-and-report'sab.shdoes this for you.
Gotchas
- Corpus grain. One call fans out to many utterance rows. Count
DISTINCT parent_interaction_idfor anything you would call a "number of conversations" — counting rows over-counts by chat volume. search.fields(GET /search/fields) lists the filter vocabulary before you write typed predicates, so you are not guessing field names.- MCP sessions can drop mid-task.
ground.shuses REST with retries, so a single blip does not lose the pull.