| Takeaway | Detail |
|---|---|
| A grounded self-assessment agent must cite work artifacts before any accomplishment is committed to the review. | Reader rule: verify the live, complete option before committing; compare like-for-like totals and terms. |
| Unaided recall is tested against grounded self-assessment on two measures: coverage and specificity of self-written accomplishments. | Controlled test design from the thesis: grounded self-assessment vs. unaided recall for performance review prep. |
| Self-assessment writing requires specific examples, not general claims, to track performance accomplishments. | Grounding: 'Tracking Performance Accomplishments and Writing Self Assessments' instructs writers to use specific examples. |
| A complete self-assessment reflects achievements, challenges, and goals for a clear and balanced view of performance. | Grounding: filling a self-assessment form involves reflecting on achievements, challenges, and goals. |
This guide tests an LLM agent over your work artifacts for performance review prep, comparing grounded self-assessment against unaided recall.
It measures coverage and specificity of self-written accomplishments, with a verify-before-you-commit rule and like-for-like comparisons.
How It Works
An LLM agent over your work artifacts works as a retrieval-and-drafting loop, not as a memory test. You point it at the raw material you already produced during the review period — commit messages, ticket comments, design docs, meeting notes, shipped pull requests, dashboards, support threads — and it indexes that corpus, then answers a single prompt: "List this person's accomplishments with evidence." The mechanism has three moving parts. First, ingestion: the agent chunks and embeds each artifact so every claim it later makes can be traced back to a specific file, line, or timestamp. Second, grounding: when it drafts an accomplishment, it must cite the artifact that supports it, which is what separates a grounded self-assessment from unaided recall. Third, reconciliation: it compares its grounded list against what you wrote from memory alone, and the gap between the two lists is the actual output you care about.
Two terms carry the whole method. Grounded self-assessment means every accomplishment statement is anchored to a retrievable artifact — the agent can show you the commit, the doc, or the ticket that proves the work happened. Unaided recall means you write your accomplishments from memory with no artifact lookup, the way most people fill out a self-assessment form. The controlled test is simply running both and diffing them. Coverage is the share of grounded accomplishments that also appear in your recall list; specificity is whether each recalled item names what you did, how you did it, and the result, which is the three-part standard self-evaluation guides consistently recommend for accomplishment statements.
| Term | What it means | How you check it |
|---|---|---|
| Grounded self-assessment | Each claim cites a work artifact | Ask the agent to show the source for every bullet |
| Unaided recall | You write from memory, no lookups | Draft your list before opening any artifact |
| Coverage | Overlap between the two lists | Count items in both, divide by grounded total |
| Specificity | What / how / result per item | Flag any bullet missing one of the three |
The verify-before-you-commit step is where the mechanism earns its keep. Before you paste anything into the review form, run the agent's grounded list against the live, complete artifact set — not a partial export, not last quarter's folder. Confirm the corpus is current and whole, because a stale index will undercount your coverage and make your recall look better than it is. Then compare like-for-like: same period, same scope, same definition of "accomplishment" on both sides. If the grounded list has twelve items and your recall has seven, the five you missed are your real prep list, and each one should already carry its citation.
One rule keeps the test honest: the agent drafts, you verify. Treat every generated bullet as a claim to be checked against its cited artifact, not as finished text. If a bullet has no traceable source, cut it or go find the artifact. The output of this section's mechanism is not a polished self-assessment — it is a verified gap list plus evidence, which is exactly what you need before you commit anything to the form.
Key Factors to Consider
When you decide whether to run an LLM agent over your work artifacts before writing your self-assessment, three criteria should drive the call: coverage, specificity, and verifiability. Coverage asks whether the grounded draft surfaces accomplishments across the full review period, not just the last two months your memory favors. Specificity asks whether each statement names what you did, how you did it, and the result — the exact structure that self-evaluation guides like ThriveSparrow's 180+ example library recommend for accomplishment statements. Verifiability asks whether every claim traces back to an artifact you actually produced, so nothing in your draft rests on recollection alone.
The numbers that matter here are few but decisive. First, count the distinct accomplishments your unaided recall produces in a timed draft, then count what the artifact-grounded pass adds; the delta is your coverage gap, and it's the single most persuasive figure you can bring to the decision. Second, check the ratio of statements that include a concrete result versus statements that only describe activity — self-appraisal collections such as Vantage Circle's 120+ examples exist precisely because result-bearing statements are the harder, higher-value form. Third, if you work in the federal system, check whether your agency operates under a rating cap such as the 40% high-performance ceiling reported by Federal News Network; in a forced-curve environment, vague self-assessments cost you disproportionately, because the margin between rating tiers is thin.
Apply a like-for-like rule when you compare the two paths. Run the unaided draft and the agent-assisted draft against the same review period, the same job duties, and the same accomplishment prompt — sample self-assessment forms, such as the free templates catalogued by SampleForms, typically ask you to reflect on achievements, challenges, and goals, so hold both drafts to that identical structure. Comparing drafts written under different prompts or different time windows tells you nothing about whether the agent added value.
Before you commit to the grounded draft, run one verification pass: for every accomplishment statement, confirm you can point to a specific artifact — a ticket comment, a design doc, a shipped deliverable — that supports it. The USDA's guidance on tracking performance accomplishments emphasizes using specific examples, and the practical test is simple: if you cannot produce the underlying artifact on request, downgrade the statement or cut it. A draft where every line survives that check is the one worth submitting.
| Decision criterion | Check to run | What the result tells you |
|---|---|---|
| Coverage | Count accomplishments in unaided draft vs. grounded draft over the same period | The delta quantifies what recall alone misses |
| Specificity | Share of statements with a named result, not just activity | Higher share means stronger, example-driven writing |
| Verifiability | Can you produce a supporting artifact for each claim? | Unverifiable lines get cut before submission |
One final framing check: self-assessments serve both you and your manager, as YourDictionary's self-evaluation guidance notes — they are a communication tool, not just a reflection exercise. That means the decision criteria above are not about producing more text; they are about producing a draft whose every claim is covered, specific, and verifiable, so the conversation with your manager starts from evidence rather than assertion.
Common Mistakes
Two mistakes sink most grounded self-assessment runs, and both are easy to catch before you commit the draft to your review packet. The first is treating the agent's output as a finished document rather than a candidate list. Suppose you point the agent at a quarter of commit messages, ticket comments, and design docs, and it returns a set of bullet points. The temptation is to paste all of them into the form. Instead, open your artifact set and count how many of those bullets trace back to a specific commit, ticket, or doc you can name. Any bullet with no traceable source is unaided recall wearing a grounded costume — the agent filled gaps from its training distribution, not from your work. The check is mechanical: for each bullet, name the artifact. No artifact, no bullet.
The second pitfall is comparing a grounded draft against your memory instead of against a like-for-like baseline. You remember the big launch and the two visible wins; the grounded draft surfaces the migration you quietly finished, the on-call rotation you stabilized, and the review comments that unblocked a teammate. If you judge the draft by whether it matches what you already remembered, you will discard exactly the accomplishments the agent was supposed to recover. Run the comparison properly: write your unaided list first, freeze it, then run the agent, then diff the two lists. The diff — not the draft — is the deliverable.
A third, quieter error is letting coverage masquerade as specificity. A bullet like "improved system reliability" covers ground but specifies nothing. The self-evaluation guidance from ThriveSparrow is blunt about the fix: every statement should cover what you did, how you did it, and the result. Apply that three-part test to each bullet the agent returns. If a bullet cannot answer all three from your artifacts, it is a heading, not an accomplishment — send it back to the agent with the specific artifact attached and ask for the how and the result.
Finally, watch for scope drift across the review period. An agent retrieving over a year of artifacts will happily pull in work from outside the window, or merge two separate efforts into one inflated bullet. Before you commit, verify the date range on every artifact you cite and confirm each bullet maps to one effort, not two. The rule is simple: one bullet, one artifact, one date inside the period.
| Pitfall | Check before committing |
|---|---|
| Untraceable bullets | Name the artifact for each bullet; drop any with no source |
| Memory as baseline | Freeze your unaided list first, then diff against the draft |
| Coverage without specificity | Each bullet must state what, how, and result |
| Scope drift | One bullet, one artifact, one date inside the period |
Insider Tactics
The non-obvious move is to stop asking the agent to write and start asking it to attack. Hand it the bullets you already believe are true, and require that it return, for each one, the raw artifact that supports it — the ticket thread, the review comment, the doc revision. Bullets that come back unanchored get cut or rewritten down to what the artifacts actually show. The rule for your final packet: every line maps to a file, thread, or document you can open on request in the room. That constraint buys more credibility than any amount of polishing.
The next move is to demand quotations rather than summaries. Ask for the exact artifact text each claim rests on, and read that text before you accept the claim. A shallow quote — a ticket title, a commit subject line — tells you the available sources is thin; a quote that names what changed and what it affected tells you the line would survive your manager's reading. ThriveSparrow's self-evaluation examples set the bar at what you did, how you did it, and the result. Apply that same bar to the agent's quotes, not just to your final prose.
Keep the wide-net job and the wording job apart. The wide-net pass should produce a flat list of candidate events with no persuasive language at all; the wording pass comes later and handles phrasing only. Blending them is how fluency hides gaps, because finished paragraphs make a short list feel complete. Your check: if a pass hands you polished sentences before you have agreed on the list underneath them, treat that list as unverified and rerun it flat.
Timing runs counter to instinct here. Write your unaided list first, before you read a single generated line, and write it in one sitting. Once you have read fluent draft prose, you can no longer tell which accomplishments you would have recalled on your own, and the comparison you are trying to make stops meaning anything. Sampleforms' guidance to reflect on job duties, achievements, challenges, and goals is a serviceable prompt for that blind pass — do it from memory, then let the agent run.
Then time the grounded run to the end of the artifact stream, not the middle. Work from the final stretch of the period lands after everything else and is often the most visible material, so a run that stops early quietly undercounts it. Run the full pass once the stream goes quiet, then a short delta pass against anything that arrived since, immediately before you submit. Hold the terms identical both times — same window, same artifacts, same list format — or your totals will not be comparable.
Comparison
Put the two options on the same table and the winner is clear: for a self-assessment you will actually have to defend in a review conversation, the grounded draft — an LLM agent reading your commits, tickets, design docs, and meeting notes — beats unaided recall on the two dimensions that decide ratings, coverage and specificity. Where a rating cap applies, such as the 40% high-performance ceiling reported by Federal News Network, every accomplishment you fail to surface is one that cannot compete for a slot in that capped pool.
| Dimension | Grounded (LLM over artifacts) | Unaided recall (memory only) |
|---|---|---|
| Accomplishments surfaced | Higher — retrieved from raw artifacts | Lower — limited to what you remember |
| Specificity (what / how / result) | Artifact carries the detail | Depends on memory fidelity |
| Verifiability | Traceable to source items | Harder to substantiate |
| Risk of omission | Lower | Higher |
Grounded wins when the review period was long, the work was cross-functional, or the artifacts live across many systems. The structure the best examples follow — "what you did, how you did it, and the result" (thrivesparrow.com) — is exactly what memory corrupts first. If your instructions tell you to "use specific examples," a draft reconstructed from tickets and commit messages gives you named, dated evidence instead of reconstruction.
Unaided recall is defensible only when the scope is small and self-contained: a short cycle, a single visible project, a handful of deliverables you authored start to finish. Even then, apply the like-for-like check — count the accomplishments you can name unaided, then count what the artifacts produce, over the same period and against the same rubric terms. When the artifact total exceeds your recall total, you have measured your own omission gap.
One caution before you commit either way: a grounded draft is only as complete as the artifact set you feed it. Verify the live, complete option — the full period, every repository, the current rubric — before investing in either approach, and compare the two outputs in identical terms (count of accomplishments, specificity per item, verifiability per claim). That comparison, not the tool, tells you whether the retrieval step paid for itself.
What to do next
| Step | Action | Why it matters |
|---|---|---|
| 1 | Before writing a single accomplishment, open the work artifacts the agent will cite — tickets, docs, shipped changes — and confirm each one is live and complete, not a draft or a stale copy. | The canonical rule is to verify the live, complete option before committing; an accomplishment backed by a dead or partial artifact collapses under review scrutiny. |
| 2 | Run the grounded self-assessment pass first, forcing the agent to attach a specific example to every claim, then run the unaided recall pass separately on the same review period. | The controlled test design from the thesis isolates grounding as the variable, so coverage and specificity can be compared without recall contaminating the grounded run. |
| 3 | Score both passes on the two measures — coverage and specificity of self-written accomplishments — using the same rubric for each. | Comparing like-for-like totals and terms is the only way to know whether grounding actually moved the numbers or just changed the phrasing. |
| 4 | Reject any accomplishment that states a general claim instead of a specific example, and send it back for a concrete artifact citation. | Grounding guidance in "Tracking Performance Accomplishments and Writing Self Assessments" requires specific examples; general claims are exactly what unaided recall produces and what reviewers discount. |
| 5 | Assemble the final self-assessment so it reflects achievements, challenges, and goals together, with each achievement still tied to its cited artifact. | A complete self-assessment needs all three elements for a clear and defensible review narrative; dropping challenges or goals leaves the document lopsided. |
| 6 | Re-verify every cited artifact one last time immediately before committing the self-assessment to the review. | Verification is a pre-commit gate, not a one-time check — artifacts can change between drafting and submission, and the rule applies at the moment of commitment. |
Frequently Asked Questions
What must an agent do before an accomplishment is committed to my review?
A grounded self-assessment agent must cite work artifacts before any accomplishment is committed to the review.
What two measures are used to compare grounded self-assessment against unaided recall?
The comparison is tested on coverage and specificity of self-written accomplishments.
What should a complete self-assessment cover beyond achievements?
A complete self-assessment reflects achievements, challenges, and goals for a clear and balanced view of performance.
Why does the grounding source insist on specific examples in self-assessments?
'Tracking Performance Accomplishments and Writing Self Assessments' instructs writers to use specific examples rather than general claims to track performance accomplishments.
What rule should I follow before committing to an option or comparing results?
Verify the live, complete option before committing, and compare like-for-like totals and terms.
How does the agent work over my work artifacts?
An LLM agent over your work artifacts works as a retrieval-and-drafting loop, not as a memory test, pointed at the raw material you already produced during the review period.
Quick answers
| What must a grounded self-assessment agent cite before any accomplishment is committed to the review? | It must cite work artifacts before any accomplishment is committed to the review. |
| What two measures are used to test unaided recall against grounded self-assessment? | Coverage and specificity of self-written accomplishments. |
| What does the reader rule say to verify before committing? | Verify the live, complete option before committing and compare like-for-like totals and terms. |
| What does the grounding guide instruct writers to use when tracking performance accomplishments? | It instructs writers to use specific examples, not general claims. |
| How does an LLM agent over your work artifacts work? | It works as a retrieval-and-drafting loop, not as a memory test, pointed at raw material you already produced during the review period. |
Also worth reading: Sales meeting prep agent: 30 to 8 minutes vs chatbot in 2026: Sales meeting prep agent: 30 · Why your AI productivity agent needs access to your chat history: Why your AI productivity agent · Let an AI agent handle your weekly priorities—no manual tracking needed: Let an AI agent handle