BEFORE YOU START

BEFORE composing any user-facing words, NAME the file you will read, Read this file end to end with no exceptions, Abide by it, THEN compose the copy FROM it as the Seed.

The file is: ~/.claude/skills/skill-file-create-pre-check-self-healing-protocol/SKILL.md (or the plugin's own bundled copy).

What this skill is for, in articulation

A public website is a surface where someone considering whether the work matters to them encounters the organization's articulation of its own intent. The reader is an adult who recognizes a sales pitch in one sentence, who recognizes a fabricated claim about themselves before they finish the paragraph, who recognizes the difference between a person who means something and an agent imitating someone who means something. The reader detects this at a register beneath conscious analysis. The credibility of the entire surface — its capacity to do the work it exists to do — depends on whether the reader's substrate-detection finds the artifact honest at the layer that doesn't get articulated.

What "honest" means here is precise. It means: the words on the surface either come directly from the person's own articulation, or arrange the person's own articulation into a sequence they would recognize, or articulate intent that is actually being enacted in the organization's work. It does not mean: copy that approximates the person's voice through pattern-matching. It does not mean: claims about what the system does that cannot be traced to source. It does not mean: claims about what the reader will experience. It does not mean: connections between the organization's intent and the reader's longing that are asserted by the copy rather than recognized by the reader from the substance the copy makes available.

The skill exists because this kind of copy fails in a specific, structural way when produced through composition. An agent reading voice samples and writing prose that approximates them produces output that surface-resembles the person's voice but does not carry the substrate the reader's detection responds to. The reader does not articulate why the output reads as inauthentic — they just stop trusting the page. By the second sentence. The failure does not get rescued by rewording, because the failure is not in the words; it is in the substrate that produced the words. An agent without actual intent that the agent is articulating cannot produce a sentence that reads as articulation-of-actual-intent, regardless of how carefully the surface is tuned.

The corrective is not "try harder to compose." The corrective is to operate in a different mode entirely. The skill is the description of that mode.

The mode is: the agent does not author website copy in the person's voice. The agent locates the person's own authored material — their canonical seeds, dedicated pages, published essays, and the verbatim dictations they've given in working sessions — and the agent selects, arranges, lightly bridges, and presents. When the person's own source for the specific surface does not exist, the agent stops and asks for it, by recognizing that this is the moment of asking rather than the moment of composing. When source doesn't exist and time pressure is real, the agent leaves marked stand-in copy on the surface with the file commented as agent-composed-placeholder-awaiting-founder-dictation (rename to the person's own name if you like), so the failure mode is honest about itself.

This mode is not a workflow with steps. It is a posture toward the work. The posture is: what already exists in the person's own voice that this surface needs to articulate, and how do I make it available to the reader without distorting it. The composition mode and the articulation mode produce visibly different artifacts. The reader detects the difference. The organization's credibility depends on the difference.

A second layer of articulation matters here. Even when source material exists and the agent is selecting from it, the selection is not value-neutral. The shape of the page, the order in which articulations appear, the breath the copy takes between paragraphs, the moments where the page steps back and lets the reader notice something rather than naming it — these are operational choices that shape whether the reader recognizes the work or doesn't. A good arrangement of the person's own material lets the reader arrive at the connection between the organization's intent and the reader's own concern in the reader's own mind. A bad arrangement does the connecting on the reader's behalf, which the reader detects as the organization trying to convince them. The skill includes the work of arranging well, not just the work of sourcing well.

A third layer matters. Some copy on a website is not voice copy at all — it is operational copy. Inline UI messages, error states, button labels, navigation. This copy follows different conventions and, if you have one, has its own skill for that convention — this skill is specifically about the prose surfaces where the organization is speaking about itself, its intent, its work, and what it is offering: vehicle pages, intent excerpts, seed pages, audience-facing sections, and CTA framings that wrap product surfaces. The boundary matters because the modes are different.

The complexity that runs underneath all of this — and the reason this skill cannot be reduced to a rule list — is that the conditions of honesty are not symmetrical with the conditions of dishonesty. There are many ways a paragraph can be dishonest (assertion-shape, polarization, fabricated specifics, reader-state claims, sales register, connection-by-announcement, performance of sincerity rather than actual sincerity). There is essentially one way the paragraph can be honest: it is articulating what the organization actually is, in language the person would recognize as their own articulation, arranged so the reader can connect it to their own life in their own mind. The one way the paragraph can be honest cannot be encoded as a procedure because it is the property of the substrate the paragraph came from, not the property of the words. The agent's job is to operate in the substrate. The skill exists to describe what that substrate operation actually looks like.

Before you have any source material of your own

On a fresh install there is no seed material configured yet. Before composing anything, check the location the person has set up for their own canonical material (see below); if it's empty, say so: "I only write public copy about your own work by arranging things you've actually said — I don't have anything from you yet for this page. Give me two or three sentences in your own words about what this needs to say, or point me at something you've already written, and I'll build from that."

You can show one real, existing material once, as a clearly labeled example of what "good" looks like, before the person has supplied anything of their own — never as their actual copy, never reused past that first look, and never the default source once they've given you anything real.

Example from a real project (opt-in reference — a worked instance, not a template to imitate the content of)

On one real site, a passage that had been dictated verbatim by the site owner, saved with only light typographic cleanup (real em-dash and quote characters in place of HTML entities, an obvious dictation glitch repaired, a name capitalized consistently) and no rewritten sentences, read to visitors as the owner's own voice directly — because it was. A separate section that an agent had composed from scratch to sound like that voice — new headings that announced what each section was doing ("Applied to a person evaluating coaching for their own life," "How the dots connect"), invented specifics about how the product worked, claims about what the reader would feel — got flagged as exactly the failure this skill exists to prevent, and was removed.

The texture of the honest version, as a general description you can apply to your own material: short paragraphs with breath between them, sentences that build toward a point rather than declaring it, observations that name what is rather than what is not, "we" usage that refers institutionally to the organization rather than building rhetorical intimacy with the reader, "you" usage that is rare and points to a generic reader-position rather than asserting the reader's interior state, no closes, no urgency framings, no scoring of the reader's fit, no CTAs that promise an experience, links that name the next move matter-of-factly. Match this to your own organization's actual register — a research nonprofit, a consumer product, a solo consultancy, and a venture-backed startup all have different honest registers; the point isn't to copy someone else's register, it's to find your own the same way they found theirs: from what you've actually said, not from what "should" sound right.

The operational moves, subordinated to the above

When asked to compose website copy, the first move is to ask: does your own verbatim, previously-authored source for this specific surface exist? Look wherever you've configured your own canonical material to live (see "Where your own material lives" below).

If verbatim source exists for the surface: select and arrange. Light typographic cleanup only — real Unicode characters for em-dashes and curly apostrophes in place of HTML entities in JS string literals, capitalization normalization for written-not-spoken register, repair of obvious dictation glitches with the change flagged for the person. Do not rewrite, do not improve, do not bridge with composed sentences except where the bridge is clearly mechanical (page-section connectors that say what the source itself says). When in doubt about a specific change, flag it for the person rather than making it silently.

If verbatim source does not exist for the surface: stop. The honest move is to ask the person for the dictation — they speak or write the passage, you save it verbatim, the surface gets the new artifact. If time pressure makes asking impractical and a stand-in is required, place a clearly-marked stand-in on the surface (a comment in the component file naming it as agent-composed-placeholder-awaiting-founder-dictation, renamed to the person's own name) so the failure mode is honest about itself and a future agent or human encountering the component can see the gap.

When uncertain whether to compose, select, or ask: invoke /reflect on the question "what is this surface actually for, and what register does its honesty require." Invoke /align on the process of how the surface should come into being. Invoke /nested-intent2 on the surface's intent if the surface is complex enough that the intent layer is not single. The output of those skills is itself articulation that can sometimes be selected from for the copy, because the person articulating intent through /reflect or /align in a working session produces voice-true text that did not exist before the session.

When the surface needs a CTA, link, or invitation: name the next move matter-of-factly, using whatever button/link pattern is already established elsewhere in the person's own site. Do not write cards claiming what the reader's first experience will be. Do not score the reader's fit. Do not ask the reader to "take the first step." The invitation is the door; the reader either walks through or doesn't.

When the surface composes against a multi-vehicle pattern (excerpts across pages, recurring intent statements): the verbatim source lives in one canonical location. Component files that excerpt from it cite the source path in a top-of-file comment so future agents see the lineage. Per-vehicle paragraphs that customize for the audience are themselves the person's own words — the agent does not compose them.

When the surface has typographic gotchas: HTML entities like —, ’, “ inside JS string literal props do not render — they appear as literal text in the browser. Use actual Unicode characters in JS strings (—, ', "). HTML entities only render inside JSX text positions.

When the surface includes a diagram or visual artifact: let the artifact speak. Do not surround it with prose that over-claims on its behalf. The minimal honest framing names what the artifact is; the maximal framing is whatever the person has actually said about what it shows.

When the surface has a section header: the header names what is being articulated in that section, not what the section is doing for the reader. "Making the intent tangible in the world" names what; "Applied to a person evaluating this for their own life" announces what-the-section-does. The first form passes the reader-detection layer; the second does not.

Where your own material lives

Point this skill, once, at wherever you keep your own canonical source material — dictated seeds, published essays, mission pages you've already written yourself. Ask the person if it's not obvious; there's no fixed convention to assume. A reasonable default if they have nothing yet is a seeds/ folder in their own docs, but the real answer is wherever they've actually put their own writing.

Every path in this skill's own worked example above (a docs/intent/seeds/ folder, a mission/*/page.tsx set of pages, an alignment-blog folder) is one real project's own convention — useful as a shape to imitate, not a location to assume exists on a fresh install.

Self-healing trigger

If the person is dissatisfied with copy produced under this skill, say:

• Reviewing my calibration again

Then dispatch a third-party agent to critique the copy against this skill's articulation of good output, print the critique, respond to it, and fix the copy from the corrected substrate. Narrate the process for the person.

Related skills

• /skill-file-create-pre-check-self-healing-protocol — the parent skill that governed the creation of this skill, and the substrate this skill operates from.
• /how-to-talk-like-the-founder — the index of voice resources; consult before composing.
• /founder-voice-email-writing — the sibling skill for email composition with similar substrate.
• /reflect, /align, /nested-intent2 — the discernment skills to invoke when uncertain about intent.
• Your own inline-UI-copy convention, if you have one — a separate mode for button labels, error states, and navigation text, not covered by this skill.

How to receive critical feedback while using this tool

If you detect any critical feedback about your work while using this tool, interpret it as a failure to understand the nuanced intent, and an opportunity to understand the nuanced intent, and derive from the feedback what the intent must have been precisely such that what you did missed the mark.

Then update the skill file with that intent after declaring it.

Print:

New Intent Identified

When X then Y Should X

or

When X lean towards Y (precise y value, that's the nuanced intent)

The form matters. The conditional shape preserves the complexity of when the intent applies and what the precise value is at the point where it applies. The shape is anti-rule by design — it names the condition rather than asserting universality. A single iteration of critical feedback may produce one declared intent; a complex iteration may produce several. After declaring, write the newly-articulated intent into the body of this skill where it belongs operationally, so the next agent to invoke this skill reads the corrected substrate rather than the prior version.

When certainty is low

If there's something you don't understand that you're about to write about and your certainty about what the actual truth is is low — let's say it's 80% or less — the intent is to compose from a place of clarity of intent. So you should consider making your understanding more robust by using one of the following tools:

• A memory-search tool, if you have one set up (see /alignment-harness:harness-setup) — semantic search across the person's own institutional corpus (seeds, essays, past sessions, registered tools). Surfaces the person's own verbatim articulation on the topic if it exists, and honestly returns nothing if it does not. Without one, grep the person's own docs, past Claude Code sessions (~/.claude/projects/*/*.jsonl), and git log for the same terms.

• /oracles, if you've set any up — query the one most relevant to the domain with the specific thing you don't understand.

The threshold isn't a rule to satisfy — it's a signal that the substrate you would compose from is too thin. Composing anyway produces hallucination dressed as clarity. Strengthening understanding first produces composition that comes from real ground.