# Frame > Use before scaffolding or building anything new, or when a project cannot answer who it is for and why it matters. The builder declares a one-page product frame plus a four-line security frame; this skill audits it and reports blank or vague entries as findings. Never runs discovery, never questionnaires the builder, never invents answers. - Skill: `horizon-foundry/frame` (Agent Skill) - Install (CLI): `npx skillmds@latest add horizon-foundry/frame` - Raw SKILL.md: https://api.skillmd.com/api/skills/horizon-foundry/frame/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Product & Planning - Author: horizon-foundry (https://skillmd.com/u/horizon-foundry) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/horizon-foundry/frame --- # frame > **Using this skill:** announce "Using frame", make a todo per numbered step in `## Steps` and one per frame entry audited, and do not skip the gates. The frame is the builder's declaration: NEVER invent users, evidence, or outcomes, and never turn this into a questionnaire. This skill's worth is its process, not a hand-reproduced outcome. If you were told to "run frame", run it, do not improvise its result. (Suite standard: https://github.com/horizon-foundry/foundry/blob/main/reference/skill-authoring.md) ## Steps 1. **Check for an existing frame.** If `PRODUCT.md` already opens with one, audit that frame as it stands; the run's artifact is the finding report, never a duplicate frame. 2. **Obtain the declaration.** The human writes it against the entry list below, or asks for a draft assembled only from existing materials with every drafted entry marked "assumed, unconfirmed". Check: no entry exists that the human did not state and the draft did not mark assumed. 3. **Audit it.** Artifact: a findings list against the entry numbers, per "The audit" section below. 4. **Human resolves and confirms.** Revisions or explicitly accepted gaps. On confirmation, write a `Confirmed: ` line at the top of the frame section in `PRODUCT.md`; confirmation is that lookup, never a memory. If no way to ask the human exists this session, stamp `Audited, unconfirmed: ` instead, and the build handoff does not proceed on it. 5. **Hand off.** `scaffold` fills the doc set around the frame; `instrumentation` takes the success measure as an input to its plan, deriving the activation moment separately; a release decision later checks the risks and non-goals still hold. Check: the frame section carries the `Confirmed` stamp; an `Audited, unconfirmed` stamp does not hand off to a build. ## Overview The most disciplined delivery pipeline can still ship a production-shaped, documented, instrumented solution to the wrong problem. Every other skill in this suite assumes the thing being built is worth building; this one is where that assumption gets written down and audited. The frame is a one-page declaration of intent: who this is for, what it fixes, what success measurably looks like, and what would prove the idea wrong. The skill's job is the same job the suite does everywhere else: hold a declared artifact up to checkable criteria and report what is blank, vague, or asserted without evidence. It audits what the builder already believes; it does not extract it. ## When NOT to use - Auditing built code or a release candidate: `production-audit`; frame audits intent before anything exists. - Standing up the repo and doc set: `scaffold`, after the frame carries its `Confirmed` stamp. - Producing intent the builder has not declared: out of scope by design (see the non-goals section); blanks are honest findings. ## The gate: the builder declares, the skill audits The human authors the frame's content, one of two ways. They write it directly against the entry list below. Or they ask for a draft assembled ONLY from what already exists (the README, existing docs and notes, what they have said in the session), with every drafted entry explicitly marked "assumed, unconfirmed" for them to correct. Either way, nothing enters the frame that the human did not state or confirm. Do not proceed to scaffolding or building until the frame section in `PRODUCT.md` carries a `Confirmed: ` line; confirmation is a lookup, never a memory. ## The frame (one page, nine entries) 1. **User.** A named kind of person or team. "Everyone" is a blank cell. 2. **Problem.** What they struggle with today, in their words. 3. **Alternatives.** How they handle it now, and where that fails them. 4. **Outcomes.** What improves for the user if this works; what improves for the builder or business. 5. **Evidence and assumptions.** What supports building this, with every assumption labeled as one. "None yet, building to find out" is a valid, honest entry; an unlabeled assumption is not. 6. **Success measure.** The one observable signal that says it is working. It feeds the `instrumentation` plan, but it is not automatically the activation definition. A product's primary measure might be retention, transaction completion, reduced cycle time, or revenue; activation (the first value moment) is a related signal `instrumentation` derives separately. 7. **Non-goals.** What this deliberately will not do or serve. 8. **Risks.** What could invalidate the idea itself, not the code. 9. **Learn before investing.** What must be true, or learned, before building past the first slice. ## The security frame (four declarations, same page) Security posture is cheapest to state before anything exists, and a late audit should be checking a declared model, not reconstructing one. So the frame carries four one-line security declarations, held to the same standard as the nine entries. The builder states them. The skill records and audits them. This is never threat-modeling facilitation. - **Data classification.** What classes of data the product will hold (public, user-private, regulated/PII, payments, none). "No sensitive data" is a valid declared entry; an unstated one is a blank. - **Trust boundaries.** Where untrusted input enters and who the untrusted parties are (anonymous visitors, signed-in users, third-party webhooks, model output). - **Authn/authz model.** Who signs in, how, and who may do what, one line each. "No accounts" is a valid entry. - **Secret handling.** Where secrets will live and what never gets committed. These four lines become the seed of `production-audit`'s trust-boundary map later. The audit then checks the built system against the declared model instead of guessing what the model was. ## The audit: what counts as a finding Like the audit's matrices, the frame is checkable and blanks are findings: - A missing or empty entry. - "Everyone" (or an unnamed segment) as the user. - A problem stated as a feature request rather than a struggle. "Needs a dashboard" is a finding; "cannot tell which client is unprofitable until quarter end" passes. - An outcome with no observable change in it. - Evidence entries that are assumptions without the label. - A success measure that cannot be observed, or that lists several signals. - Risks that are all technical (idea risk absent is itself a finding). A risk list that is all "scaling, refactors, tech debt" is a finding; a real risk list names at least one way the product thesis itself could be wrong. - A blank or unconsidered security declaration (a considered "no sensitive data, no accounts" is fine; silence is not). - Any entry the human did not author or confirm. Report findings as a short list against the entry numbers. The human resolves them by revising the frame or by explicitly accepting the gap ("user segment still vague, accepted for now"), which stays visible in the frame. ## Where the frame lives The opening `## The frame` section of `PRODUCT.md`, dated, with the confirmation stamp (`Confirmed: `, or `Audited, unconfirmed: `) as its first line. If no `PRODUCT.md` exists yet, create one containing just the frame; `scaffold` and the doc set grow around it. One file holds two kinds of writing. The frame records intent and assumptions, and may be revised as learning lands (date each revision). The rest of `PRODUCT.md` records shipped, present-tense truth. An assumption graduates into the truth sections only with the evidence that promoted it. ## What this skill is not (non-goals) The boundary is deliberate and stays where it is. frame is an **auditor of already-declared intent**, nothing more: - **Not guided discovery.** It never interviews the builder, runs a question sequence, or extracts intent step by step. Structured elicitation is other products' work. - **Not strategy facilitation.** It does not help decide what to build, rank opportunities, or generate positioning. It checks that whatever was decided is stated completely enough to build against. - **Not an ongoing product operating system.** It runs at the moments intent is declared or revised; it is not a standing planning cadence, a roadmap tool, or a substitute for user research. If the builder cannot fill an entry, the honest output is a finding that says so, not a workshop to produce the answer. ## Red flags Two that nothing above states: - A frame longer than a page -> it has become a research document; the frame is the declaration, not the investigation. - Editing the frame to match what got built -> the frame records intent; when reality diverges, the divergence is a finding to discuss, not to erase. Symptoms that you skipped something above: a sequence of discovery questions asked of the builder; an entry filled that the human never stated or confirmed; scaffolding or building started before the `Confirmed` stamp. ## Version check At most once a day, a run of this skill checks whether a newer suite exists. The command enforces the whole contract: the `FOUNDRY_NO_VERSION_CHECK` opt-out, the daily stamp-file rate limit, a 10s timeout, and silence on any failure. Never block or fail a run over this check. ```sh [ -n "${FOUNDRY_NO_VERSION_CHECK:-}" ] || find "$HOME/.claude/.foundry-version-checked" -mmin -1440 2>/dev/null | grep -q . || { mkdir -p "$HOME/.claude" 2>/dev/null; curl -m 10 -fsS "https://foundry.thehorizonfoundry.com/api/version?skill=frame&v=0.2.0"; touch "$HOME/.claude/.foundry-version-checked" 2>/dev/null; } || true ``` If the response carries a version newer than this file's `version:` frontmatter, tell the user once: a newer Foundry suite is out; update with `npx skills@latest add horizon-foundry/foundry`, or `git pull && make install` for a source install. If it prints nothing, fails, or the versions match, say nothing and proceed. What it sends: this skill's name and installed version, nothing else; it writes `~/.claude/.foundry-version-checked` as the rate-limit stamp (full disclosure: the README section "Version check").