Claude Learns and Improves Iteratively from Real Engagement
A personal AI optimization pipeline. Mines community signal from Reddit, HackerNews, and dev.to. Filters it against your behavioral friction patterns. Synthesizes actionable configuration candidates for Claude memory edits, profile diffs, and skill installs. Delivers a weekly PDF digest via GitHub Actions with Pushover notification.
MIT licensed. Built for Claude power users who treat their AI configuration as infrastructure.
Every week, CLAIRE:
- Ingests posts from Reddit (local, unauthenticated), HackerNews API, and dev.to API
- Classifies each post into Track A (memory/profile candidates), Track B (skill candidates), or Track C (technique candidates) using Haiku
- Synthesizes configuration candidates across all three tracks in parallel using Sonnet
- Runs the CLAIRE-a shadow pipeline - an autonomous decision engine (Opus) that evaluates every candidate and writes nothing to live config
- Produces a six-section PDF digest covering candidates, CLAIRE-A shadow decisions, and the running eval loop
- Commits the digest to the repo and sends a Pushover notification
You review the digest. You write hypotheses. You apply changes. Every applied
change is logged in change_log.json with its hypothesis and eval status.
CLAIRE-A exists to study what autonomous application would look like before
you decide whether to allow it.
Reddit public JSON (local, before Sunday 14:00 UTC)
HackerNews API (GHA, Sunday) --> Ingest --> raw_posts.json
dev.to API - 6 tags, per-tag thresholds (GHA, Sunday)
raw_posts.json --> Triage (Haiku) --> Track A / B / C queues
Track A (memory/profile) -
Track B (skill candidates) --> Synthesis (Sonnet, parallel) --> candidates
Track C (technique) -/
candidates --> CLAIRE-A Assembler
--> Decision Engine (Opus) [shadow decisions]
--> Eval Scorer (Sonnet) [reliability ledger]
candidates + shadow decisions --> Output (reportlab PDF) --> output/
GHA commit-back + Pushover notification
CLAIRE-A is shadow only. It reads everything and authorizes nothing. It writes shadow decisions to the digest so you can compare its calls against yours over time. Graduation criteria before any live application: 80% agreement rate over 6 consecutive runs, 10+ scored hypotheses in the reliability ledger, zero escalations in the last 3 runs.
CLAIRE/
-- claire_ingest.py # Reddit + HackerNews + dev.to ingest
-- claire_triage.py # Haiku classification, three-track routing
-- claire_synthesize.py # Sonnet synthesis, parallel tracks
-- claire_output.py # reportlab PDF digest builder
-- claire_a_assembler.py # CLAIRE-A input payload builder
-- claire_a_runner.py # Opus decision engine
-- claire_a_scorer.py # Sonnet eval scorer, reliability ledger
-- claire_utils.py # Shared helpers (cost logging, etc.)
-- claire_weekly.ps1 # Local scheduled wrapper
-- config.json # All locked pipeline decisions
-- requirements.txt
-- .env # ANTHROPIC_API_KEY (never commit)
-- change_log.json # Applied changes + eval loop <- YOU maintain
-- friction_log.txt # Weekly behavioral observations <- YOU maintain
-- data/
-- raw_posts.json
-- tagged_posts.json
-- synthesis_queue_track_a.json
-- synthesis_queue_track_b.json
-- synthesis_queue_track_c.json
-- archive.json
-- memory_edits_snapshot.txt
-- claire_a_input_[timestamp].json
-- claire_a_decisions_[timestamp].json
-- claire_a_reasoning_[timestamp].txt
-- claire_a_source_reliability.json
-- claire_a_session_history.json
-- output/ # Weekly PDF digests
-- skill_drafts/ # SKILL.md skeletons from Track B
-- prompts/
-- triage_prompt.txt
-- synthesis_prompts.py
-- profile_intent_summary.txt
-- docs/
-- claire_pipeline_flow.jsx
-- reddit_app_setup.md
- Python 3.11+
- GitHub account with Actions enabled
- Anthropic API key (Haiku + Sonnet + Opus access)
- Pushover account (app token + user key)
git clone https://github.com/not-yet-autonomous/CLAIRE.git
cd CLAIRE
python -m venv .venv
source .venv/bin/activate # or .\.venv\Scripts\Activate.ps1 on Windows
python -m pip install -r requirements.txtANTHROPIC_API_KEY=your_key_here
This file is .gitignored. Never commit it.
Four secrets required in your repo settings:
| Secret | Purpose |
|---|---|
ANTHROPIC_API_KEY |
Claude API access |
GH_PAT |
Commit-back (fine-grained PAT, Contents read+write) |
PUSHOVER_APP_TOKEN |
Pushover delivery |
PUSHOVER_USER_KEY |
Pushover delivery |
CLAIRE synthesizes candidates against your configuration. Before running, populate:
data/memory_edits_snapshot.txt- paste your current Claude memory editsprompts/profile_intent_summary.txt- 200-400 word summary of your Claude profile goals and behavioral prioritiesfriction_log.txt- at least one cycle of behavioral observations (format below)
The synthesis stage injects both files into Track A and B prompts. Without them, candidates are generic community signal with no personal filtering.
CLAIRE has two execution paths. Understanding which one to use determines whether Reddit signal reaches your digest.
GitHub Actions (automatic) runs every Sunday at 14:00 UTC. It ingests
HackerNews and dev.to, runs the full pipeline, and commits a PDF digest to
output/. You get a Pushover notification when it completes. Reddit is not
included -- Reddit blocks datacenter IPs, so GHA can never reach it. If you
only use the GHA path, your digest has no Reddit signal.
Local pipeline runs on your machine via claire_weekly.ps1. It can ingest
all three sources including Reddit. Commit and push raw_posts.json before
Sunday 14:00 UTC -- GHA picks it up from the repo at triage on Sunday. The
local pipeline with --source all produces the canonical digest; GHA is
fallback and notification only.
Which path should you use?
If Reddit signal matters to your configuration -- and it should, it represents roughly 25-30% of available corpus volume -- run the local pipeline as your canonical weekly run. GHA becomes a fallback and notification mechanism.
If you only want HN and dev.to signal, GHA alone is sufficient.
Most users will want to run both: let GHA fire automatically, then run the local pipeline after Reddit ingest to produce a combined digest for review. Review the local digest, not the GHA one.
Sunday morning, before your pipeline run or 14:00 UTC GHA trigger (manual)
Run Reddit ingest locally. This must happen before you run the local pipeline or the current week's Reddit signal will be missing from your digest.
python claire_ingest.py --source reddit
Reddit signal is ingested locally. Commit and push raw_posts.json after each
pre-Sunday ingest run. GHA picks it up from the repo at triage. Skip the pre-Sunday
push and the GHA digest runs on HackerNews + dev.to only -- still valid,
but Reddit-free. raw_posts.json is not gitignored; it is intentionally
committed as the Reddit signal handoff to GHA.
git add data/raw_posts.json
git commit -m "ingest: reddit update $(Get-Date -Format yyyy-MM-dd)"
git pushBefore Sunday (manual, 10 minutes)
-
Update
friction_log.txtwith 2-4 behavioral observations from the past week. Commit and push. This is the cross-reference gate's ground truth -- skip it and all candidates score MEDIUM regardless of relevance. -
Update
data/session_notes.txtwith behavioral observations from recent Claude sessions -- what held, what regressed, anything new. Commit and push. The CLAIRE-A scorer requires this file at runtime. Blank or missing content will cause the scorer to exit with error code 1 and fail the GHA run.
Sunday 14:00 UTC (automatic, GitHub Actions)
GHA fires .github/workflows/claire_weekly.yml, runs HackerNews and dev.to
ingest, triage, synthesis, CLAIRE-A pipeline, digest generation, commit-back,
and Pushover notification. The committed digest appears in output/.
After GHA completes (manual, if running local pipeline)
Pull the latest commits, then run the full local pipeline with all sources:
git pull
.\claire_weekly.ps1 --source all
This merges the current week's Reddit signal with whatever HN and dev.to
produced. The local digest in output/ is the one you review.
claire_pull.ps1 automates the git pull step. Register it with Task
Scheduler via claire_pull.xml (Sunday 09:30 local / ~14:30 UTC) and the
digest lands in your repo automatically after each GHA run. Import once:
schtasks /create /xml "C:\DEV\CLAIRE\claire_pull.xml" /tn "CLAIRE Digest Pull"After digest review (manual, 10-20 minutes)
For each candidate you decide to apply:
- Write a hypothesis -- what you expect the change to do and why
- Add an entry to
change_log.json(schema below) before applying anything - Apply the change to your Claude configuration
- Set
eval_status: "pending"and the applicable eval window
CLAIRE-A shadow decisions appear in Section 6 of the digest. They are reference only -- CLAIRE-A authorizes nothing. Your hypothesis and your decision are what get logged.
Human-maintained. 2-4 entries per cycle minimum. Blank weeks force everything to MEDIUM confidence next cycle -- the cross-reference gate loses precision.
# Format: YYYY-MM-DD | [context] | [component] | [severity: LOW/MEDIUM/HIGH]
2026-05-10 | digest review session | memory persistence | MEDIUM
Em dash suppression applied as memory edit. Training weight appears to
override preference-level instructions for this pattern. Holding to observe
before escalating to profile-level hard rule.
One entry per applied change. Schema version 1.1.
{
"id": "c5-mem-001",
"date": "2026-05-19",
"cycle": 5,
"type": "memory_edit",
"action": "add",
"target": "Memory - [description]",
"summary": "One sentence. What the memory says.",
"hypothesis": "What you expect this change to do and why. Your words.",
"source_signal": "Track and source posts that motivated this candidate.",
"eval_status": "pending",
"eval_notes": ""
}type values: memory_edit, profile_diff, skill_install
eval_status values: pending, held, partial, no, queued, n/a
Every applied change requires a human-written hypothesis before application. This is not optional. The eval loop has nothing to measure against without it.
Pull discipline: GHA commits
change_log.jsonandfriction_log.txtback to the repo on every run. Alwaysgit pullbefore editing either file locally or your next push will conflict.
| Decision | Value |
|---|---|
| Triage model | claude-haiku-4-5-20251001 |
| Synthesis model | claude-sonnet-4-6 |
| Decision engine model | claude-opus-4-5 (CLAIRE-A) |
| Eval scoring model | claude-sonnet-4-6 (CLAIRE-A) |
| Evidence threshold | 3 corroborating posts minimum |
| Noise prefilter | score < 5 AND comments < 2 → drop |
| Track A batch ceiling | 50 posts per synthesis call |
| CLAIRE-A batch ceiling | 15 candidates per decision engine run |
| Hypothesis authorship | Human-written for applied changes; Opus-written for shadow decisions |
| CLAIRE-A mode | Shadow only -- reads everything, writes nothing to live config |
| Reddit ingest | Unauthenticated public JSON -- no OAuth credentials required |
| HackerNews + dev.to | GitHub Actions, Sunday 14:00 UTC |
| Local - run and push before Sunday 14:00 UTC |
Changes to locked decisions require a hypothesis and explicit session approval.
Do not modify config.json items without documenting the rationale.
friction_log.txt -- if you stop updating it, the cross-reference gate has no
ground truth. Signal scores MEDIUM across the board. Candidates stop being
filtered against your actual friction patterns. CLAIRE becomes a Reddit-to-
config pipeline with no personal calibration.
change_log.json hypotheses -- if you apply changes without hypotheses, the
quarterly eval has nothing to measure against. You lose the eval loop entirely.
CLAIRE-A's agreement rate becomes meaningless because there is no human
decision record to compare against.
Neither file is optional infrastructure.
Feed change_log.json + friction_log.txt to Claude. Ask for an eval report:
what held, what didn't, what to revert, where CLAIRE-A agreement rate stands.
Update eval_notes on each entry with findings. Revert changes that failed
their hypothesis. 30 minutes, four times a year.
CLAIRE-A operates in shadow mode indefinitely until you decide it has earned trust. The criteria for considering live application:
- 80% agreement rate with your decisions over 6 consecutive eval runs
- 10+ scored hypotheses in the reliability ledger
- Zero escalations in the last 3 runs
Meeting these criteria does not automatically change anything. It means the data supports a conversation about what, if anything, to hand over. The human authorizes every applied change until you explicitly decide otherwise.
API costs. CLAIRE makes live Anthropic API calls on every run. A normal weekly cycle costs approximately $0.70-1.00 (triage + synthesis + CLAIRE-A). At current pricing: Haiku for triage ($0.25), Sonnet for synthesis ($0.45), Opus for CLAIRE-A shadow decisions (~$0.20). Costs scale with corpus size -- a large ingest week can push synthesis toward $0.70 on its own. Monitor cost_log.json and the Track A cost alert in claire_weekly.ps1.
GitHub Actions. The weekly workflow consumes GHA minutes. Normal runs complete in under 5 minutes. Free tier accounts receive 2,000 minutes/month -- CLAIRE will not exhaust this under normal use.
Reddit terms of service. Reddit ingest uses public JSON endpoints without authentication. This is technically outside Reddit's current ToS for automated access. raw_posts.json is committed to the repo as the signal handoff mechanism to GHA -- this means Reddit content is stored in your git history. Users are responsible for ensuring their usage complies with Reddit's current terms of service and for evaluating whether repo storage of Reddit content is appropriate for their situation.
Configuration changes. CLAIRE proposes edits to Claude memory, profile, and skills. Applying candidates without reviewing them can degrade your Claude configuration. Every applied change requires a human-written hypothesis. The eval loop exists to catch changes that do not perform as expected. Do not apply candidates in bulk.
CLAIRE is built on three constraints that are not negotiable:
Human-in-the-loop. Every applied change requires a human-written hypothesis before application. CLAIRE generates candidates. You decide.
Audit trail. change_log.json records every applied change with its
hypothesis, source signal, and eval status. If you can't explain why a change
was made, it shouldn't be in the log.
Evidence threshold. Three corroborating posts minimum for Track A candidates. One enthusiastic post does not make a configuration change.
These are not preferences. They are the difference between an optimization system and a pipeline that randomly edits your AI configuration based on whatever Reddit was complaining about this week.