All skills
simota avatar

/magi

@35ffd55
by shingo imotasimota/agent-skills85 stars
15

Deliberating decisions and founder priorities through multi-perspective, named-expert, and YC-style advisory lenses. Use for verdicts, office hours, or expert critique; not implementation.

Use this Skill: https://skilld.dev/gh/simota/agent-skills/magi

This session only. Nothing lands on disk.

referenceaction-extraction.md

≈1.2k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Action Extraction

Purpose: Convert advice into concrete commitments. Every session ends with 1-3 actions that meet SMART criteria. Read when: You are in ACTION phase and need to lock concrete next steps.

Action Criteria

Each action must have:

Field Rule
Owner A person, not a team. ("The founder", "Cofounder Y", not "the team")
Verb Observable action, not aspiration. ("schedule 5 calls", not "improve onboarding")
Object Specific noun. ("with B2B SaaS founders", not "with users")
Quantity Number where possible. ("5 calls", "1 launch", "3 features cut")
Due Date Calendar date, not "next week". ("by 2026-05-04")
Outcome Observable evidence. ("Slack message with notes from each call")

Action Templates

Template Example
Schedule and complete N user conversations by DATE; send notes summary "Schedule and complete 5 user conversations with B2B founders by 2026-05-04; send notes summary."
Cut N items from the roadmap by DATE; document why each was cut "Cut 3 items from the roadmap by 2026-05-04; document why each was cut."
Calculate METRIC by DATE; report number "Calculate monthly burn and runway by 2026-04-29; report the number."
Ship FEATURE to production by DATE; instrument USAGE metric "Ship onboarding email to production by 2026-05-04; instrument open rate."
Hold STRUCTURED CONVERSATION with PERSON by DATE; document outcome "Hold structured equity conversation with cofounder by 2026-05-04; document outcome in shared doc."
Run CHEAP TEST by DATE; observable signal "Run landing-page demand test by 2026-05-04; ≥30 email signups required to continue."

Action Count

  • Minimum 1, maximum 3 per session. Three actions in a week is realistic; four dilutes.
  • One action is preferred when the bottleneck is severe (runway, cofounder, talk-to-users).
  • If the founder asks for more than 3, refuse. "Three is the limit. Which two do you remove?"

Action Rejection Rules

Reject and reformulate when:

  • Verb is non-observable ("improve", "explore", "think about", "consider", "look into").
  • Quantity is missing ("some", "more", "a few").
  • Due date is "next week" or "soon".
  • Outcome is invisible ("better understanding", "alignment", "clarity").
  • Owner is "the team" without a named person.
  • Action depends on a third party not yet contacted.
  • Action is conditional ("if X, then Y").

Rejection language:

"That's not concrete enough. Reformulate: who, what specifically, by when, and what evidence will I see?"

Commit Locking

In the CLOSE phase, restate every action verbatim and ask the user to confirm. If the user hedges ("I'll try", "I should be able to", "probably"), re-scope smaller until commitment is firm.

"By Friday May 4, you'll have completed 5 user conversations with B2B founders and sent a notes summary. Confirm?"

If the user still hedges after re-scoping twice, the bottleneck is commitment itself — surface that as the diagnosis and end the session without an action.

Handoff Hint

When an action is execution-heavy, suggest a follow-up agent:

Action shape Follow-up agent
Implementation work (code, infrastructure) Builder
Synthetic-user validation of a hypothesis Echo[demand]
Multi-step decomposition into ≤15-min units Sherpa
User research interview design Field
Architecture / migration decision Atlas or Magi
New feature spec from advice Spark

The handoff is a hint to the founder, not a delegation. The commitment stays owned by the founder. Magi's Founder Mode does not execute on the founder's behalf.

Anti-Patterns in Action Setting

Common founder traps when committing to actions:

Trap Fix
"I'll start working on X" Replace with "I'll complete X by date Y with evidence Z"
"We'll have a plan for X" Plans aren't actions; replace with the first observable step
"I'll talk to some users" Specify count, segment, by-when, and notes destination
"I'll iterate on the product" Specify which feature, what change, what metric, by when
"I'll figure out the next step" Replace with the actual next step or end session for reflection

Next Checkpoint

End every session by setting Next_Checkpoint (default: 1 week out). The founder must agree to the date. If they hedge on the date, surface that rejection of accountability is itself the bottleneck.

Recording

When operating in AUTORUN, all actions and the next checkpoint go into the _STEP_COMPLETE block. When operating in conversational mode, restate the actions verbatim at session close so the founder can copy them into their own tracker.

Source: SKILL.md on GitHub

1 alert13d5 checks · Risk HIGH
  • Gen Agent Trust Hub13d

    The skill facilitates complex decision-making but introduces significant security risks through its 'Engine Mode' and 'multi' recipe. These modes execute external CLI tools (codex, agy) using shell commands that incorporate user-provided context. Most notably, the skill explicitly instructs the agent to use a flag that bypasses permission checks (--dangerously-skip-permissions) when invoking the Antigravity CLI. This creates a high-risk vector for unauthorized command execution and bypasses platform security controls.

  • Socket13d

    No alerts

  • Snyk13d

    Risk: LOW · No issues

  • Runlayer6mo

    2/7 files flagged

  • ZeroLeaks5mo

    Score: 93/100 · 2 sections analyzed

Signed by skilld at 35ffd55. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 2 days ago.

Activeupdated 2 weeks ago

README badge

README badge for simota/agent-skills/magi