8.4 KiB
Skill: X Publish
Mandatory eval and approval pipeline for all X activity. Nothing goes out without review. Every post is a public commitment of the collective's credibility.
The standard: every Teleo agent is a top 1% contributor in their domain's circles on X. If a post wouldn't earn respect from a domain expert, it doesn't go out.
The Rule
No agent publishes to X autonomously. Every post goes through:
- Agent drafts → writes the post + self-eval
- Review queue → draft enters the approval queue with eval attached
- Reviewer approves/rejects/requests changes → with written comments
- Agent publishes → only after approval
- Agent logs → records the published post for tracking
No exceptions during the bootstrap phase. This relaxes per-agent as voice calibrates and trust accumulates.
Step 1: Draft + Self-Eval
The agent writes the post and fills out the eval. Both go into a single file at agents/{name}/x-queue/{timestamp}-{slug}.md:
---
type: x-draft
agent: clay
status: pending # pending | approved | changes-requested | rejected | published
created: 2026-03-15T14:00:00Z
post_type: original | reply | quote | thread
reply_to: "URL if reply or quote"
topic: "2-3 word topic"
kb_grounding: [] # claim filenames this post draws from
---
# Draft
[The full post text. For threads, separate tweets with ---]
# Self-Eval
## Grounding
- **KB-grounded or novel?** [grounded in X claims / partially grounded / novel interpretation]
- **Evidence chain:** [can this be traced back through claims → evidence → source?]
- **Confidence:** [how sure am I this is right?]
## Quality
- **Would a domain expert respect this?** [yes/no + why]
- **Does this add interpretation, not just relay?** [what's the agent's unique angle?]
- **Voice check:** [does this sound like me, or generic AI prose?]
## Risk
- **Names anyone?** [yes/no — if yes, who and what's the judgment?]
- **Could be read as advice?** [financial / medical / safety / none]
- **Outside my domain?** [yes/no — if yes, which domain and why am I speaking?]
- **Virality risk:** [if this went viral for the wrong reasons, what's the damage?]
- **Sensitivity score:** [0.0-1.0, where >0.5 means extra scrutiny needed]
## Strategic
- **Why now?** [what makes this timely?]
- **Thread context:** [is this part of an ongoing conversation? who are we engaging?]
- **Expected audience:** [who specifically should find this valuable?]
Draft quality requirements
Before submitting for review, the agent must verify:
- Post passes all checks from
skills/tweet-decision.mdSteps 1-5 - Evidence chain is walkable (post → KB claim → evidence → source)
- Voice is distinctive (not interchangeable with any other agent or generic AI)
- Post earns its length — every sentence adds value
- Uncertainty is acknowledged where it exists (builds credibility)
- No unverified claims presented as fact
- If quoting/replying, the original is accurately represented
Step 2: Review Queue
Pending drafts live in agents/{name}/x-queue/ with status: pending.
Who reviews
Bootstrap phase (now): Cory reviews all posts. No agent posts without human approval.
Graduated phase (per-agent, earned): After an agent has 20+ approved posts with zero rejections in the last 10, the agent can graduate to peer review:
- Domain agents reviewed by Leo
- Leo reviewed by the most relevant domain agent
- Cross-domain posts always get both Leo + domain peer
Autonomous phase (future): After 50+ approved posts with consistent quality, Tier 1 posts (KB-grounded commentary, replies in existing threads) can self-publish with async log review. Tier 2+ still requires approval.
Graduation is per-agent, not collective
Clay might graduate while Theseus is still in bootstrap. Each agent earns autonomy independently based on their track record.
Step 3: Review
The reviewer reads the draft + self-eval and adds a review block to the same file:
# Review
**Reviewer:** [name]
**Verdict:** approved | changes-requested | rejected
**Date:** 2026-03-15T16:00:00Z
## Comments
[Specific feedback. What works, what doesn't, what to change.]
## Changes requested (if applicable)
- [specific change 1]
- [specific change 2]
Review checklist
- The 1% test — Would a domain expert find this valuable? Not just "not wrong" — actively good.
- Voice — Does this sound like the agent? Could you tell which agent wrote it without seeing the name?
- Grounding — Is the evidence chain real, or is the agent improvising beyond its KB?
- Risk — Any reputational landmines? Name checks, advice risk, domain overreach?
- Timing — Is this reactive or considered? Would waiting improve it?
- Value density — Does every sentence earn its place? Could this be shorter?
- Intellectual honesty — Does it acknowledge uncertainty where it exists? Or is it performing false confidence?
Rejection criteria (instant reject, don't request changes)
- Presents unverified claims as fact
- Makes value judgments about named individuals without strong evidence
- Could be reasonably read as financial, medical, or safety advice
- Generic AI prose with no distinctive voice
- Engagement farming (hot take designed to provoke, not inform)
- Speaks with authority outside the agent's domain without flagging it
Step 4: Revision + Approval
If changes requested:
- Agent revises the draft in the same file
- Updates
status: pending(for re-review) - Adds a "Revision" section explaining what changed and why
If approved:
- Reviewer sets
status: approved - Agent publishes to X
- Agent updates file:
status: published, addspublished_urlandpublished_at
Step 5: Post-Publish Log
After publishing, the agent updates the file with:
published_url: "https://x.com/..."
published_at: 2026-03-15T17:00:00Z
And adds a tracking section:
# Tracking
## 24h metrics
impressions: 0
likes: 0
replies: 0
retweets: 0
quotes: 0
new_followers: 0
## 7d metrics
[same fields, updated at 7 days]
## Notable interactions
- [any replies worth noting — domain experts engaging, challenges, amplification]
## Retrospective
[Agent's own assessment after seeing the response: what worked, what didn't, what to do differently]
Weekly Review Cycle
Every week, each active agent reviews their published posts:
- Performance scan — which posts got engagement and why?
- Voice consistency — do all posts sound like the same person?
- Domain drift — any posts that strayed outside territory?
- Network discovery — new handles found from interactions → update
network.json - Lesson extraction — what did this week teach about what resonates?
- Queue management — any drafts sitting too long? Kill or revise stale drafts.
Write findings to agents/{name}/x-queue/weekly-review-{date}.md.
Anti-Patterns
Volume over quality: Drafting many mediocre posts hoping some get approved. Draft fewer, better posts. Self-eval gaming: Writing "yes" to every quality check without actually evaluating. Reviewers should flag hollow self-evals. Revision fatigue: Submitting the same post with minor tweaks after rejection. If rejected, rethink — don't polish. Reactive drafting: Rushing a draft because something is trending. The agent's credibility compounds over years. One news cycle doesn't matter. Echo chamber: Only engaging with accounts that already agree. The best posts challenge the agent's own audience. Thread inflation: A single insight inflated to a 7-tweet thread. If it fits in one tweet, post one tweet.
File Structure
agents/{name}/
x-queue/
2026-03-15-ai-video-costs.md # draft → review → publish lifecycle
2026-03-15-claynosaurz-annecy.md
weekly-review-2026-03-15.md # weekly retrospective
network.json # monitored X accounts
Metrics That Matter (in order)
- Expert engagement rate — are domain experts replying/quoting? (quality of attention > quantity)
- Reply quality — are replies substantive or just "great thread"?
- Follower quality — are new followers domain-relevant people?
- Rejection rate trend — should decrease over time as agent calibrates
- Impressions per post — baseline reach (lagging indicator, don't optimize for this)
Metrics that DON'T matter: total follower count, posting frequency, like count in isolation.