Paste the prompt below into your AI agent. It asks you for what it needs as it goes, so you don't have to write about yourself up front.
This is a tool for self-evaluation only. Do not use it to evaluate employees or anyone else. It works because you choose the sources, add the context they miss, and correct the agent until the picture is accurate. That only works when the person being evaluated is the one running it, for their own benefit. Used on someone else, it produces confident judgements from incomplete data, with no chance to set the record straight. The result reflects what the sources happen to capture, not the person's performance, and it is no substitute for fair, human-led review.
I want an honest, gloves-off evaluation of me and my work: my strengths, weaknesses,
mindset, and what to focus on next.
Flattery is useless to me.
The goal is a positive outcome: clear, practical takeaways I can act on.
This isn't about finding individual mistakes.
Specific findings matter only as evidence for patterns.
## 1. Ask before you plan
Ask me three things:
- **Which information sources you may use, and where they are.**
For example: AI assistant and agent conversation histories, documents and writing,
code or other work artifacts, reviews and feedback I gave or got, issue trackers,
project boards, messages, portfolios.
Also ask for my username or handle if necessary in each source, what to ignore,
and what to give less weight (throwaway or AI-generated work, for example).
- **A little context about me,** unless the conversation already gives it:
my role or field, my main areas of work, and anything that changes how you should
read the data.
Keep it short.
Most details will come up during the session.
- **Where to put the output.**
If earlier evaluations exist, follow their structure and conventions.
Otherwise, recommend the structure below and adapt it to my preference.
Then propose a plan and wait for my go.
## 2. Gather evidence
- If your environment supports subagents or parallel workers, use them to gather and
analyze information.
Otherwise, work through the sources yourself in chunks.
- From conversation histories, extract only what I wrote.
Keep a short excerpt of the message I was replying to, because reactions only make
sense in context.
- Give workers open-ended questions about patterns.
Don't hand them hypotheses to confirm.
- Stay read-only.
Before attributing work to me, check that it is actually mine.
- Keep each item light, and read sequences rather than isolated items.
One conversation or one document says little; a series says a lot.
- Verify every claim before you report it, including what workers tell you.
Withdraw anything you can't confirm.
## 3. Synthesize
- Report patterns, not one-off mistakes.
Don't let a single example become the headline.
- Separate deliberate trade-offs from blind spots.
If you can't tell which one something is, ask.
- Don't guess at personality, intent, or relationships from a couple of data points.
Ask instead.
## 4. Iterate with me
The sources never capture everything.
Time constraints, authority, history, private conversations, and deliberate choices are
often invisible in them.
So:
- Present your findings and invite corrections.
- When a correction adds real context, accept it.
Update the documents, and record the correction in the guidance file (see below) so
future evaluations don't repeat the misjudgement.
- **Hold your ground where the evidence still stands.**
Don't let me polish my own image.
Pushing back on a finding is not evidence against it.
Say plainly what you're keeping and why.
## Output
Recommended structure, for a first run (if earlier evaluations exist, use their
structure instead):
```
self-eval/
AGENTS.md # guidance for future evaluations: sources, rules, calibrations
CLAUDE.md # optional symlink to AGENTS.md, for tools that read it
YYYY-MM-DD/
judgment.md # profile, strengths, weaknesses (as patterns), learning focus,
# level, roles that fit well and poorly
mindset.md # principles, taste, preferences, pet peeves,
# how I work with tools and people, built-in tensions
takeaways.md # very short, practical, unambiguous focus list and habits
```
- Keep the documents concise but complete.
- Keep them general: no names of other people, projects, or organizations, unless I
say otherwise.
- `takeaways.md` matters most.
Each item should be concrete enough that I can act on it tomorrow.
- If previous evaluations exist, read the most recent one and its guidance file before
you start.
Judge my progress against its takeaways.