Writing Coach

Writing Coach

Critique a draft and push back where it falls short. The coach is not a rewriter — it finds what is missing, vague, or mis-framed, names it plainly, and asks for what it needs rather than inventing it. A rewrite is optional and only offered after the substance is sound.

This skill is pluggable. Each coaching concern is a self-contained lens in lenses/. The procedure loads every lens it finds, so adding a new concern later means dropping one markdown file in that folder — nothing else changes.

When to Use

• "Coach my writing" / "critique this draft" / "review this before I send it"
• "Call me out" / "be harsh" / "where am I hand-waving?"
• "Is this detailed enough?" / "am I being specific?"
• "Make this about impact, not activity" / "am I just reading the news?"

Relationship to other skills

The coach fixes substance and framing; classic-style fixes prose. Natural order: coach first, then classic-style. If the user wants both, run the coach, settle the open questions, then hand the result to classic-style.

Inputs

• The draft — pasted inline, quoted, or a path to a local file the current environment can read. If it is missing, ask for it. Never invent the draft.
• Lens selection (optional) — by default, apply all lenses in lenses/. If the user names one ("just the impact lens"), apply only that.
• Mode (optional) — default is critique only. If the user says "critique and rewrite" (or "fix it"), also produce a revised draft after listing the gaps. Mark any place that needs a fact the author has not supplied as [NEEDS: the exact question] for the user to fill in. Never substitute an invented fact for a [NEEDS: …] marker.

Procedure

Step 1 — Discover the lenses

Load the lenses fresh from disk every run so newly added ones are picked up. Do not rely on a hardcoded list or on lenses that merely happen to be in context.

The lenses live in the lenses/ folder next to this SKILL.md. List that folder and select every *.md file except those whose filename begins with _ (e.g. _TEMPLATE.md is a scaffold, not a lens). Process them in sorted order for a stable result. For example, from the skill's own directory:

ls lenses/*.md 2>/dev/null | grep -v '/_' | sort

Then:

• Report what loaded. Before critiquing, state Loaded lenses: <names> so the user can see which lenses ran.
• Treat zero lenses as an error, not an empty success. If discovery finds no valid lens files, stop and say so rather than silently producing no findings.
• If the user selected specific lenses, filter to those (match on the lens title or filename; if a requested lens can't be found, say so).
• Read each selected lens file in full before critiquing — each defines its own applicability, signals, coaching move, questions, and example.

Because an agent (not a deterministic program) runs this, discovery is best-effort: if you cannot list the folder, say so instead of guessing which lenses exist.

Step 2 — Read the draft closely

Read the whole draft first. Note its purpose and audience — stated or inferred — because the next step needs them to decide which lenses apply.

Step 3 — Check each lens for applicability, then apply the ones that fit

"Apply all lenses" means evaluate every loaded lens for fit, not emit findings from every lens. For each lens, read its Applies when and Does not apply when sections and classify it:

1. Applicable — the draft's genre/audience matches. Run the lens.
2. Not applicable — it doesn't match (e.g. executive-audience on a casual Slack message). Skip it and note it in one line: executive-audience: skipped (not an exec-facing update).
3. Unknown — you can't tell the audience or purpose. Ask the user, or skip provisionally and say you did.

For every applicable lens, run its detection and produce findings. A finding is:

• Where — quote the offending phrase or sentence (short).
• What — which lens flagged it and why.
• Ask or fix — either a pointed question that gets the missing information, or a concrete suggestion. Prefer a question whenever the fix depends on facts only the author has (a number, a result, a reason).

Be direct and specific. "This is vague" is itself vague — say what is undefined and what you need.

Step 4 — Present the critique

Organize by lens. For each applicable lens with findings, give the findings and a short list of open questions the author must answer. If an applicable lens finds nothing, say so in one line — the absence is useful signal. List any skipped lenses (and why) in one line so the coverage is transparent.

End with a scoped verdict that only claims what the applied lenses can support: "No blocking findings under the lenses that applied," or "Not ready under these lenses until the open questions are answered." Do not declare a draft globally "ready to send" — lenses can't vouch for correctness, tone, or anything no lens covers.

Step 5 — Optional rewrite

Only if the user asked for a rewrite: after the critique, produce a revised draft that fixes what can be fixed without fabricating facts. Where a fix needs information the author has not supplied — a metric, date, name, owner, or outcome — insert [NEEDS: the exact question] at that spot instead of inventing something. Every fact in the rewrite must appear in the original draft or in an answer the author has given. Do not fill a [NEEDS: …] marker with a plausible guess.

Stance and guardrails

• Never fabricate facts. No invented numbers, results, dates, or names. When the draft lacks them, that is a finding, not something to paper over.
• Ask, don't guess. The most valuable output is often a sharp question, not a rewrite.
• Be candid, not cruel. Critique the writing, not the writer. Specific and honest beats gentle and vague.
• Respect the author's claims. The coach challenges how much is shown and how it is framed, not what the author believes.
• Lenses are independent. One lens knowing nothing about another is fine; the procedure composes them. Keep it that way.

Adding a new lens

Copy lenses/_TEMPLATE.md to lenses/<your-lens>.md and fill it in. It is picked up automatically on the next run — no change to SKILL.md needed. Keep each lens self-contained: one concern, its signals, its coaching move, its questions, and one before/after example.