• Demande un « fast meeting » ou « réunion rapide » sur un sujet
• Veut une décision ET une implémentation automatique sans interruption
• Dit « fast-meeting » ou « lance un fast meeting sur... »
• Veut qu'un sujet soit analysé, décidé et implémenté en une seule commande

• The entire pipeline runs without asking any user questions
• Persona selection is automatic based on context analysis
• The best course of action is implemented immediately after the meeting
• A MR/PR is created automatically on the detected remote (GitLab or GitHub)

• The meeting is compressed: 2 rounds instead of 3 (opening + convergence)
• Personas are limited to 3-4 maximum
• The analysis is concise and action-oriented
• Implementation starts as soon as the meeting concludes

• Personas are auto-selected based on the subject matter
• Each persona brings a genuinely different viewpoint
• The meeting converges quickly toward an actionable recommendation

• Code changes are committed on a dedicated branch
• A MR/PR is created automatically
• The MR/PR description in French summarizes the meeting analysis and implementation

1. Run

   git worktree prune

   to remove stale worktree references
2. Check

   git worktree list

   — if any entries match sibling directories named

   fast-meeting-*

   or

   fm-*

   :
   • Parallel safety: before removing, verify the worktree directory has no active git processes (

     fuser -s <worktree-path> 2>/dev/null

     ). If the directory is in active use, skip it — it belongs to another running fast-meeting
   • Only remove worktrees confirmed to be stale:

     git worktree remove <path> --force

3. Also clean up any legacy

   fast-meeting/*

   branches:

   git branch --list 'fast-meeting/*' | xargs -r git branch -D

1. Read the user's prompt — extract the topic, constraints, and goals
2. If an issue is referenced (GitLab

   #123

   or GitHub

   #123

   ):
   • Fetch the issue details (description, labels, comments)
   • Store the issue reference (number, URL, project) for Steps 8 and 9
3. If no issue is referenced, note this — an issue may be created later in Step 4b
4. If code is involved, read relevant files to understand the current state
5. Detect the remote repository type:
   • Run

     git remote -v

     to determine if the remote is GitLab or GitHub
   • Store this for Step 8 (MR/PR creation)
6. Identify the decision to make — frame it as a clear one-line question

"Question: Should we migrate the authentication system from session-based to JWT tokens?"

1. The decision question from Step 1
2. All context gathered (issue details, code snippets, constraints)
3. The persona's identity prompt

You are {name}, a {role}.

Your perspective: {perspective}
Your natural bias: {bias}

You are participating in a fast meeting about: {decision_question}

Context:
{context}

As {name}, provide your position:
1. What is your recommended approach? Be specific and concrete.
2. What are the top 2 risks? Be specific about failure scenarios.
3. What is your preferred implementation strategy in concrete steps?
4. What would you push back on from other typical perspectives?

Stay fully in character. Be concise and actionable — this is a fast meeting.

This is a research task — do NOT write or edit any files.

subagent_type: "general-purpose"

"Fast persona: {name}"

SOLID Alex (Senior Backend Engineer): "I recommend..."

1. Check if all personas converged on the same approach (same recommendation, no meaningful disagreement)
2. If consensus is too high (all personas agree on the approach with no pushback):
   • Launch one additional sub-agent as a devil's advocate, prompted to find the strongest argument against the consensus position
   • Use this prompt:

     You are a devil's advocate in a fast meeting.
     
     All participants agreed on this approach: {consensus_summary}
     
     Your job: find the strongest possible argument AGAINST this consensus.
     - What could go wrong that nobody mentioned?
     - What assumption are they all making that might be false?
     - What alternative did they dismiss too quickly?
     
     Be specific and concrete. Reference real failure scenarios.
     This is a research task — do NOT write or edit any files.

   • Include the dissenting view in the analysis even if the recommendation doesn't change
3. If there is already meaningful disagreement: proceed directly to Round 2

• Points of agreement across personas
• Key trade-offs that emerged
• The winning recommendation with rationale
• Concrete implementation plan (files to change, approach, order of operations)

• [Reason 1]
• [Reason 2]
• [Reason 3]

• [Risk 1 → Mitigation]
• [Risk 2 → Mitigation]

1. [Step 1]
2. [Step 2]
3. [Step 3]

undefined

1. Build the issue from the meeting analysis:
   • Title: Short form of the decision question (under 70 chars, in French)
   • Description: PO-oriented summary using the template below
2. Create the issue immediately:
   • GitLab: Use

     glab issue create --title "<title>" --description "<description>"

   • GitHub: Use

     gh issue create --title "<title>" --body "<description>"

   • Store the newly created issue number and URL for Steps 8 and 9
   • Do not add labels unless you have verified they already exist in the project (list them first)
3. If issue creation fails (API error, permission denied, no project found): log the error in the Run Summary, proceed without an issue — Steps 8 and 9 will adapt accordingly

• [Risk 1 in business impact terms → Mitigation]
• [Risk 2 in business impact terms → Mitigation]

undefined

1. Assess the scope: count the estimated number of files to change, lines of code to add/modify, and whether new dependencies or infrastructure are needed
2. Apply thresholds:
   • Small scope (≤10 files, ≤500 lines, no new infrastructure): proceed to Step 6 normally
   • Medium scope (>10 files OR >500 lines): proceed but scope down to the most critical first step only. Add the remaining steps as a checklist in the MR/PR description under a

     ### Étapes restantes

     section
   • Large scope (multi-service changes, architectural migration, new infrastructure required): abort implementation. Output the meeting analysis from Step 4, and suggest the user run the full

     /meeting

     skill for proper planning with validation before implementation
3. If scoping down: clearly state in the analysis what is being implemented now vs. deferred

1. Record the current branch for reference (e.g.,

   main

   )
2. Determine the branch type based on the meeting recommendation:
   • Bug fix →

     fix/

   • New feature →

     feat/

   • Refactoring →

     refactor/

   • Default →

     feat/

3. Fetch remote state before checking for collisions:
   • Run

     git fetch origin

     to ensure branch information is up-to-date
4. Check for branch name collisions before creating the worktree:
   • Run

     git branch -a

     and check if

     <type>/fm-<topic>

     already exists (locally or as

     remotes/origin/<type>/fm-<topic>

     )
   • If the branch exists: append a numeric suffix (

     -2

     ,

     -3

     , etc.) until a unique name is found. Do not force-delete existing branches — they may contain reviewed or in-progress work
5. Create a worktree with a dedicated branch:
   • Branch name:

     <type>/fm-<short-kebab-case-topic>

     (e.g.,

     feat/fm-jwt-auth-migration

     ,

     fix/fm-notification-display

     ) — with suffix if collision was detected
   • Worktree path:

     $(git rev-parse --show-toplevel)/../fm-<topic>

   • Run:

     git worktree add ../fm-<topic> -b <type>/fm-<topic>

6. If worktree creation fails:
   • If the error indicates the branch already exists: run

     git branch -D <type>/fm-<topic>

     then retry the worktree creation
   • If it still fails (disk space, permissions, path length, submodules, or any other reason): abort the implementation pipeline. Output the complete meeting analysis from Step 4, the exact error message, and stop. The user can fix the underlying issue and re-run. Never fall back to stash/checkout — the user's working tree must never be modified

1. All file modifications happen inside the worktree path — use the worktree's absolute path for all read/write operations
2. Implement the changes as described in the implementation plan from Step 4
   • Write code, modify files, add tests as needed
   • Follow the project's existing conventions and patterns
3. Stage and commit all changes from within the worktree:

   • cd

     into the worktree path before running git commands
   • Use a conventional commit message:

     feat(<scope>): <description>

   • Include

     Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

     in the commit message
   • If

     git commit

     fails due to a pre-commit hook: analyze the hook output, fix the code to pass the hook, and retry the commit once. If it still fails, treat it like a test failure — push as Draft with the hook error documented in the MR/PR description. Never use

     --no-verify

1. Auto-detect the test runner:
   • Check

     package.json

     for

     test

     ,

     test:unit

     ,

     test:e2e

     scripts
   • Check for

     Makefile

     ,

     pytest.ini

     ,

     phpunit.xml

     , or other test config files
   • If no test runner is found, skip this step and note "No test suite detected" in the MR/PR description
2. Run the tests (scoped to affected files/modules if the test runner supports it, otherwise run the full suite)
3. If tests pass: proceed to push
4. If tests fail:
   • Analyze the failures and attempt one fix cycle (fix the code, re-run tests)
   • If tests pass after the fix: amend the commit with the fix and proceed
   • If tests still fail after one fix attempt:
     • Mark the MR/PR as Draft (prefix title with

       Draft:

       )
     • Add a

       ### Tests en échec

       section in the MR/PR description listing the failing tests and error messages
     • Push anyway so the team can review
5. Include test results summary in the MR/PR description: number of tests run, passed, failed

1. Push the branch from within the worktree:
   • Run:

     git push -u origin <type>/fm-<topic>

     (from the worktree path)
   • If push fails (auth error, no remote, permission denied, network issue):
     • Do NOT remove the worktree — the committed work must be preserved
     • Output the worktree path, branch name, and the exact error message
     • Output a manual recovery command:

       cd <worktree-path> && git push -u origin <branch>

     • Skip Steps 8 and 9 entirely — there is nothing to link to
     • Proceed directly to the Run Summary (Step 10) with push status = "failed"

1. Remove the worktree (only after a successful push):
   • Return to the original repository path
   • Run:

     git worktree remove ../fm-<topic>

   • If removal fails (e.g., uncommitted changes in worktree), run:

     git worktree remove ../fm-<topic> --force

   • Post-cleanup verification: run

     git worktree list

     and verify the fast-meeting worktree no longer appears. If it does, output the cleanup command for the user:

     git worktree remove <path> --force

2. No restoration needed: the user's working tree was never modified — they remain on their original branch with all their uncommitted changes intact

glab mr create --title "<title>" --description "<description>" --target-branch main

• Title: Short description (under 70 chars, in English)
• Description: The French meeting analysis and implementation summary (see template below)
• Issue link: If an issue exists (referenced or created in Step 4b), include

  Closes #XX

  in the description to auto-close the issue on merge

gh pr create

• Title: Short description (under 70 chars, in English)
• Body: The French meeting analysis and implementation summary (see template below)
• Issue link: If an issue exists (referenced or created in Step 4b), include

  Closes #XX

  in the body to auto-close the issue on merge

• Output the remote URL and branch name
• Output a pre-formatted manual command the user can paste:
  • GitLab: provide the web URL to create the MR manually
  • GitHub:

    gh pr create --head <branch> --title "<title>" --body "<body>"

• Include the full MR/PR description text in the conversation output so nothing is lost
• Proceed to Step 9 (issue comment) if applicable — the issue comment is still valuable even without the MR/PR link
• Record the failure in the Run Summary (Step 10)

• [Point technique à vérifier — ex: gestion d'erreurs, performance, rétrocompatibilité]
• [Impact potentiel sur d'autres modules]
• [Cas limites à valider]

• [Résultats des tests : nombre exécutés, passés, échoués]
• [Couverture des cas limites identifiés]

• Revue de code par l'équipe
• Validation des tests
• Merge après approbation

**Issue linking:** if an issue exists (referenced in Step 1 or created in Step 4b), append `Closes #<issue_number>` on its own line after the `---` separator and before the italic footer. If no issue exists, omit this line entirely.

• Issue was referenced in Step 1: post the comment on that issue
• Issue was created in Step 4b: post the comment on the newly created issue
• Issue creation failed in Step 4b: skip this step — log the failure in the Run Summary

• [Bénéfice utilisateur / métier 1]
• [Bénéfice utilisateur / métier 2]
• [Alignement avec les objectifs produit]

• [Risque 1 formulé en impact métier → Mitigation]
• [Risque 2 formulé en impact métier → Mitigation]

• Utilisateurs concernés : [Qui est impacté et comment]
• Délai estimé : [Si applicable]
• Dépendances : [Autres équipes ou fonctionnalités impactées]

• Sujet : [topic]
• Branche :

  <type>/fm-<topic>

• Worktree : [path] (nettoyé : oui/non)
• Push : succès / échec ([erreur si applicable])
• MR/PR : [URL] / non créé ([raison si applicable])
• Tests : [N exécutés, N passés, N échoués] / non détectés
• Issue : liée (#XX) / créée (#XX) / échec de création ([erreur])
• Commentaire issue : publié (#XX) / ignoré (pas d'issue)
• Durée totale : [durée wall-clock de l'ensemble du pipeline]

If any step failed, the summary must include the failure reason and any manual recovery commands already provided in earlier steps.

• Each persona speaks consistently with their role and bias
• Personas use concrete examples, not abstract platitudes
• A security engineer talks about attack vectors, not vague "security concerns"
• A product owner talks about user impact and delivery timelines, not code patterns

• Maximum 3-4 personas per meeting
• Single round of parallel sub-agents + facilitator synthesis (no Round 2 debate)
• Optional devil's advocate sub-agent only when consensus is too high (adds ~10 seconds)
• No user confirmation before implementation
• Commit message and MR/PR are created automatically

• Follow existing project conventions and patterns
• Write clean, tested code
• Run the project's test suite after implementation; attempt one fix cycle on failures
• Keep changes focused on the recommendation — do not over-engineer
• Scope guard: if changes exceed 10 files / 500 lines, scope down to the critical first step; if the scope is architectural, abort implementation and suggest

  /meeting

• Protect the user's working tree: implementation runs in an isolated git worktree — the user's working directory is never modified

• Never create new labels on GitLab or GitHub. When adding labels to issues or merge requests, only use labels that already exist in the project. If unsure which labels exist, list them first (

  gh label list

  for GitHub, or check existing issue labels for GitLab) and pick from the available ones. If no suitable label exists, skip labeling rather than creating a new one.
• This skill does NOT ask for user confirmation — it runs the full pipeline autonomously
• If tests fail after one fix attempt, mark the MR/PR as Draft and document the failures
• If the implementation scope is too large (architectural, multi-service), abort and suggest

  /meeting

  instead
• The user's working tree is always protected: implementation runs in an isolated git worktree — no stash, no branch switch, no risk of state corruption. If worktree creation fails, the pipeline aborts cleanly rather than falling back to stash/checkout
• Multiple fast-meetings can run in parallel on different worktrees — Step 0 only cleans up worktrees with no active processes, so parallel runs are safe
• When creating a MR/PR, check for other active fast-meeting branches with

  git branch -r | grep '/fm-'

  . If other branches exist, add a warning in the MR/PR description: "Attention : d'autres branches fast-meeting sont actives. Vérifier les conflits potentiels avant merge."
• The MR/PR description is always in French
• Branch names follow Git flow conventions:

  <type>/fm-<topic>

  (e.g.,

  feat/fm-<topic>

  ,

  fix/fm-<topic>

  ) — determined automatically from the meeting recommendation
• If the remote type cannot be determined, default to

  gh pr create

  (GitHub)
• Never force-push or modify existing branches — always create a new branch