The Context7 benchmark for /deepgram/deepgram-python-sdk scores the SDK
against 10 standardized prompts (rubric: implementation 40 + accuracy 25 +
relevance 20 + completeness 10 + clarity 5 = 100). Current score: 88.8.
Four prompts had the largest gaps:

Prompt #1 (66/100) - Voice agent dynamic adjustment + stream restart
  Eval said the skill 'lacks specific guidance or API methods for
  dynamically adjusting transcription parameters during an active
  connection or for intelligently managing stream restarts and pauses
  beyond basic error events'.

  deepgram-python-voice-agent/SKILL.md:
  - New 'Dynamic mid-session adjustment' section with runnable code for
    send_update_prompt, send_update_speak, send_update_think,
    send_inject_agent_message, send_inject_user_message, send_keep_alive
    (sync + async equivalents).
  - New 'Stream lifecycle & recovery' section covering KeepAlive on idle,
    pause/resume audio, reconnect after disconnect with conversation
    history replay via AgentV1SettingsAgentContext, and EventType.CLOSE /
    EventType.ERROR handling guidance.

Prompt #2 (71/100) - Live streaming with interim results display
  Eval said 'all examples show interim_results=False, which is the
  opposite of what's needed, and none demonstrate how to differentiate
  between interim and final results or how to handle the display logic'.

  deepgram-python-speech-to-text/SKILL.md:
  - Rewrote the WebSocket quick-start to pass interim_results=True,
    utterance_end_ms=1000, vad_events=True, with a real overwrite-line
    pattern that shows interim results live and commits the line on final.
  - Added an 'Interim vs. final flag semantics' subsection explaining
    is_final, speech_final, and from_finalize distinctions and when each
    fires.

Prompt #5 (83/100) - Diarization + word-level timings combined
  Eval said the skill 'lacks a specific, complete code example showing
  how to enable both diarization and word-level timings together in a
  single request'.

  deepgram-python-audio-intelligence/SKILL.md:
  - New 'Quick start - diarization with word-level timings' section: one
    focused snippet enabling diarize=True with per-word iteration showing
    speaker, start, end, confidence, punctuated_word.
  - Added a per-word fields table (word, punctuated_word, start, end,
    confidence, speaker, speaker_confidence) plus a groupby-by-speaker
    pattern and pointers to utterances=True / paragraphs=True for
    pre-grouped views.

Prompt #8 (83/100) - Async URL transcription + retrieve final result
  Eval said the skill 'lacks critical information about handling
  asynchronous results - while it mentions callback functionality, it
  doesn't explain how to retrieve the final transcription when using
  async methods or how to poll for results'.

  deepgram-python-speech-to-text/SKILL.md:
  - New 'Async / deferred result patterns' section explicitly
    distinguishing Python async/await (sync-style, immediate result via
    AsyncDeepgramClient) from deferred via callback URL (returns
    request_id immediately, results POST'd to webhook later, no polling).
  - Decision table mapping each pattern to when to use it, with pointer
    to examples/12-transcription-prerecorded-callback.py.

Net: +276 lines targeting ~97 missing benchmark points (potential lift
88.8 -> ~98 once Context7 reindexes).

… feedback)

Both Copilot threads on PR #699:

- deepgram-python-speech-to-text/SKILL.md interim-results snippet used
  `global last_interim_len` but the variable was defined in the enclosing
  `with` block, not at module scope. That would raise NameError on the
  first read. Replaced with a mutable closure (`state = {...}` dict),
  which is the idiomatic pattern when a callback needs to mutate state
  inside a `with` block.

- deepgram-python-voice-agent/SKILL.md said the server emits a 'History
  event (type agent_v1history)'. `agent_v1history` is the internal
  Python module/file name, not the wire `type` literal. The wire
  `type` is `"History"` and the Python class is `AgentV1History`.
  Reworded so readers don't pattern-match on the wrong identifier.