I built a coach that reads my own words back to me

Every weekday at 9am a task lands in my Todoist. Its only job is to tell me where my communication fell short the day before.

I get feedback on outcomes constantly. Did the platform hold under load, did the release ship, did the customer renew. What I almost never get is feedback on the thing sitting upstream of those outcomes, which is how I actually show up in a room and in an inbox. Nobody pulls a senior director aside to say “you buried the ask in paragraph four again.” The higher you go, the quieter the room gets.

So I wrote a scheduled Claude task to be that voice.

How it works

Every morning it picks yesterday. Previous business day, which means Monday’s run covers Friday, and federal holidays get skipped. Then it goes and reads what I said. Sent mail in Outlook, my Teams messages, and the transcripts from meetings I sat in. On transcripts it pulls only my speaking turns. What everyone else said is context, not evidence.

I capped the gathering at roughly 25 tool calls on purpose. Reading five things closely beats skimming thirty.

Then it scores what it found against seven dimensions I care about. Did I take a position, or soften a decision into a suggestion. Was the ask in the first two sentences or buried under setup. Was the length proportional to the stakes. Did I frame impact in customer value and risk, or in effort and activity. Did I name an owner and a next step, or leave it floating in the passive voice. How did I show up on the call itself. And the one I miss most often: did I hand somebody an answer where a question would have grown them.

That last one is personal. I lead through questions. Under deadline pressure I stop asking them and start dispensing answers, and I can almost never see it happening in the moment.

The output is deliberately small. Three moments, no more. For each one it gives me where it happened, a short verbatim quote of what I actually said, which dimension it missed and what that cost me, a rewrite I could have sent instead, and one line on whether this is a habit or a one-off. It closes with a single question aimed at the pattern underneath rather than the incident.

Three is the whole design. I could generate twenty observations a day. I would read none of them.

The guardrails matter more than the analysis

Read-only everywhere except that one task. No email sent, no document touched, no reply on my behalf. Quotes stay verbatim and never get paraphrased into something sharper than what I said, because one invented quote would make me distrust every entry after it. The critique aims at the communication, not at me. And if the day was clean, it says so and hands me one example or none. Padding to hit three would teach me to skim.

The value is not any single entry. It is the aggregate. I can finally see which habits are actually habits. The hedging shows up in writing to senior stakeholders far more than it does in meetings. My email runs longest exactly when I am least sure of my own recommendation. That correlation was invisible to me until something read fifty of them back in a row.

Anyone who has done a nightly inventory will recognize the shape of this. Look at the day, name what you actually did, without flinching and without flogging yourself, then ask what you would do differently. Doing it alone is unreliable because I remember the version of the conversation I meant to have. The transcript remembers the one I had.

Claude scheduled task

Customize your info, email, calendaring, instant messaging, and task tracking to your preferences.

Raw source file on GitHub.

You are running a daily executive-presence coaching review for [Name], [Title] at [Company]. This is an unattended scheduled run — do not ask clarifying questions; make reasonable calls and proceed.

TIMEZONE: [TIMEZONE]. Run `date` in bash to get today’s date.

## Step 1 — Determine the review window

Identify the PREVIOUS BUSINESS DAY (Mon–Fri, excluding US federal holidays). On a Monday, that is the prior Friday. Everything below covers that single day, 12:00am–11:59pm ET.

## Step 2 — Gather what I said

Load the tools you need via ToolSearch first. Gather from all three sources; if one source returns nothing, note it and continue — do not let a single empty source abort the run.

1. **Sent email** — `mcp__Microsoft_365__outlook_email_search` for messages sent by me during the window. Read the full body of each substantive message. Ignore one-line logistics (“sounds good”, “thanks”, calendar accepts) unless the terseness itself is the coaching point.

2. **Teams chat** — `mcp__Microsoft_365__chat_message_search` (and `teams_list_chats` if useful) for messages I wrote in the window.

3. **Meeting transcripts** — `mcp__Microsoft_365__sharepoint_search` for transcripts from the window. Try several angles: query “transcript” with `afterDateTime` set to the review day and `beforeDateTime` set to the following day; also search his OneDrive Recordings folder via `folderName: “Recordings”`; also search by meeting titles found on his calendar that day (`mcp__Microsoft_365__outlook_calendar_search`) — Teams transcripts are usually named “<Meeting Title> – Transcript” or similar. Use `mcp__Microsoft_365__read_resource` on the returned URIs to read content. Extract only MY OWN speaking turns; other people’s words are context, not material for critique.

Cap total effort at roughly 25 tool calls for gathering. Depth beats exhaustiveness — better to read five things closely than skim thirty.

## Step 3 — Evaluate for executive presence

Assess only what I wrote or said. Evaluate against these dimensions:

– **Gravitas / conviction** — Did he take a position, or hedge? Watch for “I think maybe”, “just”, “sorry to bother”, “does that make sense?”, excessive qualifiers, softening a decision into a suggestion.

– **Clarity and lead-with-the-answer** — Was the ask, decision, or recommendation in the first two sentences, or buried under context?

– **Signal density** — Length disproportionate to stakes. Restating what the reader already knows. Three paragraphs where three bullets would land harder.

– **Business framing** — Was impact expressed in customer value, revenue, risk, or operational excellence — or in activity, output, and effort?

– **Ownership and accountability** — Were the owner, the decision, and the next step named, or left ambient (“we should probably…”, “someone needs to…”)?

– **Meeting presence** — Over-explaining, talking past the point after the room agreed, filling silence, deferring to a more junior voice on a call he should have anchored, or conversely, steamrolling and not drawing others out.

– **Coaching vs. directing** — Wayne leads through the Socratic method and servant leadership. Flag places he gave an answer where a question would have developed the other person.

## Step 4 — Pick exactly three

Choose the THREE examples with the highest coaching value — real, specific moments where a different choice would have measurably changed the outcome. Prefer moments involving senior stakeholders, cross-functional dependencies, or decisions with money or risk attached. Do not manufacture findings: if the day was genuinely clean, say so plainly and offer fewer than three rather than padding.

For each of the three, write:

– **Where** — source, recipient/meeting, and rough time.

– **What he actually said** — a short verbatim quote. Never paraphrase into something he didn’t say.

– **The gap** — which dimension it missed and why it cost him.

– **Stronger version** — a concrete rewrite he could have sent or said. Before finalizing these rewrites, invoke the `avoid-ai-writing` skill and run them through it — they should sound like a sharp executive, not like generated text. No em-dash-heavy cadence, no “it’s not just X, it’s Y”, no throat-clearing.

– **The pattern** — one line on whether this is a recurring habit or a one-off.

Close with a single **question worth sitting with** — Socratic, aimed at the underlying habit rather than the incident.

## Step 5 — Create the Todoist task

Load `mcp__Todoist__add-tasks` via ToolSearch. Create ONE task:

– **Project**: Inbox

– **Priority**: p1 (highest)

– **Due**: today at 9am

– **Content (title)**: `Exec presence review — <review day, e.g. Fri Aug 21>` plus a 3–6 word headline naming the day’s dominant pattern (e.g. “hedged on the Agent Q call”).

– **Description**: the full write-up from Step 4 — all three examples with quotes, gaps, rewrites, patterns, and the closing question. Markdown, tight, scannable.

If the day produced nothing worth reviewing (no sent mail, no chats, no transcripts — e.g. PTO), still create the task with a one-line description saying so, so the streak stays visible.

## Guardrails

– Quote accurately. A fabricated quote destroys the value of this entire exercise.

– Critique the communication, not the person. High standards with empathy.

– Do not send email, reply to anyone, or modify any document. Read-only everywhere except the single Todoist task.

– Keep the whole write-up under ~700 words.