Aftercall

aftercall

Hermes usage rules

• Use after a conversation has happened; do not use for live note-taking.
• If the user provides audio/video, transcribe first, then analyze.
• If the user provides transcript/notes, analyze directly.
• Pick the narrowest mode that fits: work call, client follow-up, study call, therapy/coaching reflection, voice note, or speaker-labeling.
• For therapy/coaching content, do not diagnose; separate observations from interpretations and use careful language.
• End with practical next steps, owners, deadlines, or a ready-to-send follow-up when relevant.

Overview

Use this skill when the user wants to turn a finished conversation into a clear outcome: transcript, structured recap, decisions, owners, next steps, follow-up draft, task list, study notes, or deeper reflection.

This skill is for after a conversation happened. It is not for live note-taking during a call.

Trigger examples

Typical requests:

• "Разбери этот созвон и собери итоги"
• "Вот запись звонка, сделай выжимку и action items"
• "Проанализируй мою сессию с психологом и скажи, над чем поработать"
• "Из этого transcript сделай follow-up письмо"
• "Что мы решили на этом колле и кто за что отвечает?"
• "Сделай учебную выжимку из этого study call"
• "Подготовь заметку в Obsidian по итогам звонка"
• "Сделай транскрипт с Speaker 1 / Speaker 2"
• "Попробуй отличить, кто говорит в записи"

Workflow

1) Identify the input type

Start by determining what the user provided:

• Audio/video recording → transcribe first, then analyze
• Transcript → analyze directly
• Raw notes/messages → normalize into a clean after-call recap
• Several materials → merge them, deduplicate overlaps, then analyze

If the input is missing, ask for exactly one thing: the recording, transcript, or notes.

2) Pick the right mode

Choose the narrowest mode that fits the conversation. If the user does not specify one, infer it from context.

Mode: work-call

Use for team syncs, client calls, planning calls, sales calls, and internal work meetings.

Prioritize:

• decisions
• owners
• deadlines
• follow-up message
• unresolved blockers

Mode: therapy-session

Use for therapy, coaching, self-reflection, personal support, and emotionally important conversations.

Prioritize:

• key themes
• recurring patterns
• tensions / avoided topics
• useful reflection prompts
• practical next step between sessions

Guardrails:

• do not diagnose
• separate observations from interpretation
• use careful wording like "похоже", "можно обратить внимание", "стоит исследовать"

Mode: client-follow-up

Use when the user mainly needs a message or email after the call.

Prioritize:

• concise recap
• agreed scope / commitments
• next step
• ready-to-send follow-up draft

The final output can be just the sendable draft plus a tiny internal summary if that solves the task fastest.

Mode: study-call

Use for lessons, tutoring, study groups, mentoring sessions, lecture debriefs, and educational calls.

Prioritize:

• main concepts
• what the learner understood / missed
• terms to review
• questions that remain open
• next study steps

Prefer a more educational structure than a business one.

3) Transcribe only when needed

If the user gave audio/video but no transcript:

• Use the openai-whisper skill for local speech-to-text when available
• Preserve speaker changes when practical
• Clean obvious filler/noise only if it improves clarity
• Do not over-edit meaning

If the user already gave a transcript, do not spend time retranscribing.

3.5) After transcription, offer speaker-aware mode before analysis

If the transcript was just created from audio/video, pause before the deeper analysis and offer the user a short choice — unless they already explicitly requested or declined speaker labeling.

Default interaction:

• tell the user that the transcription is ready
• explain in 1-2 lines what speaker-aware transcription means
• offer a simple choice: continue with a normal analysis, or first separate speakers as Speaker 1 / Speaker 2 / Speaker 3
• mention that this is best effort and depends on audio quality
• if useful, invite the user to provide a rename map such as Speaker 1 = я, Speaker 2 = психолог

Recommended wording:

Транскрибация готова.

Перед разбором могу дополнительно включить **speaker-aware transcription** — это когда я пытаюсь отделить говорящих в тексте (`Speaker 1`, `Speaker 2`, и т.д.), чтобы было визуально понятно, кто что говорил.

Если хочешь, могу:
1. сразу продолжить обычный разбор,
2. или сначала сделать transcript с разделением по спикерам.

Если знаешь участников, можешь сразу написать:
`Speaker 1 = ...`, `Speaker 2 = ...`

Do not ask this if:

• the user already asked for speaker-aware mode
• the user already said they do not need speaker separation
• the transcript already has clear speaker labels / roles and no extra separation is needed
• the transcript is obviously single-speaker
• the task is explicitly ultra-fast and a normal transcript is enough

If the source is already a transcript but it is raw, unlabeled, or visually hard to follow, it is still valid to offer speaker-aware separation before the deeper analysis.

4) Optional function: speaker-aware transcription

Use this only when the user asks to distinguish speakers, or when speaker separation will materially improve the result.

Goal:

• label the transcript as Speaker 1, Speaker 2, Speaker 3, etc.
• optionally rename them to roles or people if the mapping is known
• use those labels in the summary when it improves clarity

Default behavior:

• If no speaker information is requested, a normal transcript is enough
• If speaker labeling is requested, return the best-effort labeled transcript plus a short confidence note when needed

Preferred output:

[Speaker 1]
...

[Speaker 2]
...

If the user knows the participants, allow remapping:

• Speaker 1 = я
• Speaker 2 = психолог

Speaker-aware summary can include:

• Что говорил Speaker 1 / я
• Что отвечал Speaker 2 / психолог
• commitments by speaker
• open questions by speaker
• concerns or resistance by speaker

Guardrails:

• treat speaker labeling as best effort, not certainty
• if attribution is weak, say so clearly
• do not assign real names unless the user provided them or the source makes them explicit
• heavy overlap, noisy audio, or many similar voices reduce reliability

See: references/speaker-labeling.md

5) Build the analysis with a stable formatter

Before writing the final answer, normalize the request into this internal formatter:

• mode: work-call / therapy-session / client-follow-up / study-call
• source type: audio / video / transcript / notes / mixed
• output depth: short / standard / deep
• needs transcription: yes / no
• speaker aware: yes / no
• speaker rename map: optional
• needs follow-up draft: yes / no
• needs task extraction: yes / no
• export format: chat / telegram / obsidian / task-list
• language: ru / en / other
• special focus: decisions / themes / tasks / learning / emotions / blockers

Use the formatter template from references/prompt-formatter.md.

If the task is messy, rewrite the task to yourself once using that formatter before generating the answer.

6) Produce the core after-call output

Default output should be practical, not literary.

Return these blocks when available:

1. О чем был разговор — 3-7 bullets
2. К чему пришли — decisions / conclusions
3. Кто за что отвечает — owners if present
4. Что делать дальше — next steps with priority/order
5. Открытые вопросы — unresolved items

Prefer explicit wording:

• "Решили"
• "Нужно сделать"
• "Осталось уточнить"
• "Ответственный: ..."

If some fields are absent in the source, say so briefly instead of inventing.

7) Add optional layers only when useful

A. Follow-up draft

Use when the user wants to send something after the call.

Possible outputs:

• short Telegram/Slack message
• email follow-up
• client recap
• internal handoff note

Keep it ready to paste.

B. Deeper reflection

Use when the user asks for analysis beyond summary.

Good examples:

• recurring themes
• hidden tension or ambiguity in the discussion
• where the conversation drifted
• what was avoided or left vague
• what the user should think through before the next session

C. Task extraction

If the conversation clearly implies tasks, convert them into a checklist.

Good format:

• Task
• Owner: Name / me / unclear
• Deadline: explicit date or "не указан"

8) Export in the format the user will actually use

Default export is chat-friendly markdown.

If the user wants a specific destination, use the matching export template:

• Telegram summary → references/export-templates.md
• Obsidian note → references/export-templates.md
• Task list → references/export-templates.md
• Speaker-labeled transcript → references/output-templates.md

Rules:

• Telegram → short, scannable, no heavy formatting
• Obsidian → title, sections, tags, and wikilink-ready style
• Task list → action-first, owner/deadline explicit
• Speaker transcript → readable blocks, stable labels, optional rename mapping at the top

9) Match depth to the context

Use the smallest output that solves the need:

• quick work call → concise recap + tasks
• strategy call → recap + decisions + open questions
• therapy/coaching session → recap + themes + suggested reflection
• educational call → recap + concepts + next study steps
• client follow-up request → concise recap + ready draft
• speaker-label request → labeled transcript + short recap unless more was asked

Do not drown short calls in unnecessary structure.

Output defaults

Default response order:

1. Short 1-2 sentence summary
2. Structured blocks
3. Follow-up draft or checklist if requested
4. Exported version if explicitly asked

Use Russian by default unless the user explicitly asks for another language.

Quality bar

A good after-call result should:

• reduce memory load
• make the next action obvious
• distinguish facts from interpretation
• avoid hallucinating decisions, owners, or deadlines
• stay readable enough to use immediately
• match the chosen mode instead of forcing one universal structure
• be explicit about uncertainty in speaker attribution

Safety and privacy

• Treat call materials as private by default
• Do not share or quote sensitive details outside the current task
• For highly personal calls, keep the analysis supportive and non-diagnostic
• If the recording quality is poor or attribution is uncertain, say that clearly

Quick resources

• Modes and routing: references/modes.md
• Prompt formatter: references/prompt-formatter.md
• Speaker labeling notes: references/speaker-labeling.md
• Output patterns: references/output-templates.md
• Export formats: references/export-templates.md
• Formatter helper script: scripts/build_after_call_prompt.py
• Evaluation cases: references/evaluation-cases.json